Changelog

Connector writes now open a draft for review first

Writing data into a connected system should feel like filling out a form, not firing a command. Agents now open a draft first for driver, vehicle, and account creation, so you see exactly what will be written and can change it before anything is sent.

Highlights:

• Draft first, write second: creating a driver, vehicle, or account opens a prefilled draft for review and edit; the agent is steered to draft rather than invent required values
• No stray confirmation cards: a direct write no longer spawns an empty form alongside the result, closing a path that could look like a second, unintended entry
• Smarter outbound scrubbing: the safety scrubber now recognizes operational dates on dispatch records, such as a licence or plate expiry, so a legitimate date stops being treated as a date of birth
• Steadier builds: the workspace's fonts are now bundled with the app instead of fetched at build time, removing a class of flaky deploys

The homepage now replays real recorded sessions

A demo should show the product, not a stand in for it. The homepage hero now replays three sessions recorded from real runs, an operations report, a coverage map, and a revenue forecast, each frozen to a stable export so the replay stays honest over time.

Highlights:

• Real recorded replays: the hero cycles through three actual recorded sessions instead of a staged animation
• Reports render for real: the report lightbox renders the underlying charts directly, rather than a static image standing in for them
• Maps open properly: the coverage map report opens in a new tab, working around the sandbox limits that blocked it inline
• A calmer first frame: the replay starts from its still frame and only then plays, removing an initial flash

Solutions Marketplace: modest demo pricing, and a refreshed dispatch reference

Two smaller fixes that both matter for a live demo: the Solutions Marketplace catalogue now prices at levels that make sense to try out, and the docs stopped describing a dispatch connector that had moved on without them.

Highlights:

• Modest demo pricing: catalogue prices across packs, skills, and knowledge items dropped to approachable demo tiers
• Dispatch reference regenerated: the docs' dispatch connector reference is rebuilt from the live tool contract, so read and write tools listed there match what is actually served
• Stale paths corrected: a handful of outdated fixture and path references in the docs were fixed alongside the refresh

Tenant admin management arrives in the control panel

Every workspace eventually needs someone who can manage who has access to what. That now lives in the control panel: a tenant admin can promote or change a teammate's role in a few clicks, and the change carries a full audit record of who did it and when.

Highlights:

• Role management in the control panel: a tenant admin changes a teammate's role directly, no support request needed
• Audit trail: every role change is recorded with the acting admin, closing the loop between an action and who took it

All AI on Google Cloud, plus a security pass

The AI in a HIPAA safe workspace should sit inside the same boundary as everything else. As of this release it does: every model the platform serves, without exception, runs on Google Cloud Vertex AI under the same agreement and the same controls that cover your data. A round of session and request hardening shipped with it.

Highlights:

• All AI on Vertex AI: every model the platform uses is served through Google Cloud Vertex AI, inside the compliance boundary, with no external model endpoint in the path
• Server enforced idle timeout: an idle session expires on the server, not merely in the browser tab, so a walked away from workstation stops being an open door
• CSRF protection: state changing requests carry and verify a token, closing the cross site request class of attack
• Per agent model choice: each agent can be pointed at the model that suits its job, rather than every agent sharing one setting
• Interactive tool cards: connector results can render as interactive cards you can act on in the conversation, instead of a wall of returned text

Charts that mean something, and skills that stay out of the way

Two threads landed together. Reports and dashboards can now draw genuine charts from your data instead of approximating them, and running a skill stopped taking over the conversation it runs in.

Highlights:

• A real chart kit: reports and dashboards render data driven charts, drawn from the underlying numbers rather than assembled by hand
• Skill runs as a chip: a running skill collapses into a compact chip in the transcript, so a long procedure does not bury the conversation around it
• Inline prompt fields: when a skill needs an input, you fill it in place in the chat rather than being sent somewhere else to supply it
• Insert a skill from the slash menu: type slash in the composer and pick a skill without leaving the message you are writing
• What's New in the app: a panel in the workspace lists recent releases, with an unread indicator so you notice when something new arrives

Maps go full screen, and dispatch data comes in live

Operations work is spatial, and a map squeezed into a panel is a map you argue with. Maps now open full screen, and the routing behind them answers faster. Alongside that, live operational data from the dispatch system reaches the workspace through its own connector.

Highlights:

• Full screen maps: open a map to the full window for route, corridor, and service area work, then drop back into the conversation
• Faster geo turns: route and fare quote requests come back noticeably quicker, which matters when you are iterating on a quote rather than running one
• A dispatch connector: live operational read tools appear on the connector catalog, so an agent can look up current dispatch state as part of its work
• Dashboards on the data lake path: the demo dashboards now read through the data lake, the same path production reporting uses
• Users and roles: an admin Users list lets tenant admins see who has access and manage their roles, without a support request

The lineup, simplified: connect, buy, create, run

The product family just got easier to explain, because each part now answers one question. Connect what? Nucleus. Buy what? The Solutions Marketplace. Create what? Agentic Studio. Run what? Agentic Operations. The biggest change is the marketplace: capabilities that used to be sold as large product lines are now individual items you can find, buy, and put to work the same day, priced to what each one is worth rather than bundled into something bigger than you needed.

The lineup:

• Nucleus, to connect: bring third party software onto the OneOps Gateway as a live mesh, import and harmonize and synchronize records, provision and deprovision business assets, and feed a continuous stream into your data lake. Priced per vehicle per month
• Solutions Marketplace, to buy: the one place to find and buy skills, agents, and knowledge. Dispatch product knowledge, a cross system intelligence dashboard, a broker integration agent, each an item you buy on its own. Priced per item
• Agentic Studio, to create: build and test your own skills and agents, and the place to go when the marketplace does not already have what you need
• Agentic Operations, to run: the runtime quality environment for everything you bought or built, with protections against AI harm, AI cost, and HIPAA exposure, plus logging, monitoring, undo, and an emergency stop
• Support: your own AI support robot, with human escalation available
• Consulting: time with an expert, scoped and delivered under a statement of work

What this means in practice: a skill is a discrete capability and an agent can carry several skills plus a workflow that ties them together. You can buy either one, try it in Studio, and operate it full time in Operations. Nothing you already run changes name or behavior.

AI Studio: one search box, one way in

The workspace had grown enough that finding things had become its own task. This release replaces hunting with one search box, and trims the navigation down to a single obvious way in.

Highlights:

• Universal search: a search palette opens from the header on any page and reaches across your workspace, so you stop navigating to the place where the thing you want lives
• A simpler sidebar: Projects is now the single entry point, with the per project detail where it belongs instead of duplicated in the rail
• A breadcrumb in the header: at a glance, which project and which page you are on, and one click back up
• Assisted new project setup: creating a project starts with a short guided form that gathers the essentials, rather than dropping you into an empty conversation to figure it out
• Mobile fixes: layout corrections across the small screen views, so the workspace holds together on a phone

AI Studio: project memory you approve, and a steadier composer

Projects accumulate facts worth keeping: how a client prefers to be invoiced, which depot covers which corridor, what a recurring report is supposed to include. Until now those lived in whichever conversation happened to mention them. Project memory turns them into something the project holds on purpose, with a human deciding what gets kept.

Highlights:

• Propose a memory from chat: when something in a conversation looks worth remembering, it can be proposed as a memory with a title, a short description, and the fact itself
• The owner decides: proposals arrive as a review card with Approve and Dismiss, and an approved memory lands in the project's Memory tab where anyone on the project can read and edit it
• Consolidated composer controls: the options that used to be scattered across the message box now sit in one menu, which also makes the composer usable at phone width
• The assistant dock stays put: the dock rebinds to the project chat you were actually in and survives a page reload, so a conversation you started from the dock is still there when you come back
• Cleaner spoken responses: dictation and read aloud produce natural speech, without markup or formatting artifacts leaking into what you hear

AI Studio: documents that survive the export

An export is only useful if the document that comes out the other end looks like the one you approved. This release closed the gap: tables stay tables, nested lists keep their levels, and formatted text keeps its formatting in both Word and PDF. Finished exports also stop being downloads you have to shepherd somewhere, since they now land directly in the project's Files.

Highlights:

• Export fidelity: tables, nested lists, and formatted text render correctly in both DOCX and PDF, so a document you export matches the one you reviewed
• Exports land in Files: a finished export is saved into the project's Files automatically, where the rest of the team can find it, instead of only reaching your downloads folder
• A wider Files panel: the panel opens roomier and a file opens in a viewer right in the flow, so you can read a document without leaving what you were doing
• Spreadsheets welcome: .xlsx uploads are supported alongside the existing document types
• Folder uploads: drop a whole folder into a project and its contents come across together
• A limit you can see: the 25 MiB per file cap is stated up front, and a file that exceeds it gets a clear message saying so rather than a generic failure

AI Studio: a skills marketplace you can see before you run

The skills catalog now reads like a marketplace. Skills, the reusable procedures your team captures once and runs many times, sit on department shelves you can browse instead of a flat list you have to already know your way around. The bigger change is what a skill shows you: open one and a worked example of its output renders right there, so you can judge the shape of the result before you spend a minute running it.

Highlights:

• Browse by department: the catalog groups skills onto Operations, Finance, Compliance, Sales, Marketing, Support, and Engineering shelves, plus a General shelf for cross team work, with live counts and a one click filter down to a single shelf
• See the output first: a skill's example renders inline as the real thing, a full page report at true desktop width or a live map with mapped stops, routed corridors, and a drive time area, and expands to fullscreen for a proper look. Every example is captured from an actual run on demo data, not mocked up
• The whole artifact range: skills draw on everything AI Studio produces, from interactive reports and data tables to documents that export to PDF and Word, flowcharts and diagrams, maps, slide decks, and generated images
• Run in a new chat: one button opens the skill as a fresh conversation with the procedure already loaded and the first turn already moving, no copying and pasting a prompt
• It asks rather than interrogates: when a skill needs inputs it does not have, the assistant asks for them conversationally, one at a time and explained in the skill's own terms, instead of a blank form standing between you and the work
• A first party starter set: the catalog ships with OneOps authored skills across every department, each carrying a Seed badge so it is obvious what came in the box and what your team added
• Nothing reaches a shelf unreviewed: new skills still go through the draft and approve flow, where an AI writes the draft and a workspace admin approves it, and a project can pin the shelf down to just the skills that matter to it

The OneOps lineup: Nucleus and the products on top

The OneOps product family is now named for what it does, and every surface (the platform, this site, the docs) has been swept to match. It starts with Nucleus, the foundation every customer creates first: connect your industry products, collect their data into your private data lake, and optionally keep key records like drivers, vehicles, and staff synchronized across systems. Everything else builds on that foundation.

The lineup:

• Nucleus: the mandatory foundation. Business Org, Product Connector, Product Data Lake, and Product Synchronizer connect your software, centralize your data, and keep systems in step
• Agentic Studio: the HIPAA-safe workspace for creating, experimenting with skills, agents, automations, and AI chat over your own connected data
• Agentic Operations: the HIPAA-safe runtime where your creations, and products from the marketplace, run the business day to day
• Product Intelligence: deep AI expertise on the industry products you already run, starting with dispatch platforms
• Business Intelligence: dashboards and analytics over everything in your data lake

What changed in the product:

• Product names updated across the workspace, the marketing site, and the documentation, including navigation, product pages, and guides
• Commercial catalog aligned: plans and SKUs now carry the new names while existing subscriptions carry forward unchanged
• Earlier changelog entries and page addresses updated so links and history stay consistent with the new names

AI Studio: notification center and a wave of chat reliability

Two threads landed together: the workspace now tells you what happened while you were elsewhere, and the chat itself got a deep reliability pass driven by real usage feedback. A notification bell in the app shell collects completed agent runs, questions an agent needs answered, and project archive notices into one slideover. Under the hood, a sweep of fixes targeted the failure modes that erode trust in an AI workspace: answers that loop, turns that vanish when interrupted, and tool results that say nothing.

Highlights:

• Notification center: a bell and slideover collecting completed runs, agent questions waiting on you, and archive notices, wired to producers across the platform
• No more repetition loops: a guard detects and stops degenerate repeating output mid-stream instead of letting it fill the transcript
• Interrupted turns repair themselves: stopping a response mid-stream no longer leaves a broken turn behind, and regenerating picks up cleanly
• Honest tool cards: when a tool returns nothing, the card says so and points to where long output lives instead of rendering an empty shell
• Failures stay visible: when a turn fails, a persistent inline error card explains it, rather than the message silently disappearing
• Diagrams in documents: document artifacts can now include flowcharts and diagrams that render in preview and carry through to PDF and Word export
• Live building states: maps, tables, and diagrams show a building shimmer while the assistant writes them, so long artifacts read as progress, not silence

AI Studio: skills graduate, slash commands, and a smarter composer

Skills, the reusable step-by-step procedures your team captures once and runs many times, have graduated from the Labs section into the workspace proper. They now live where the work happens: launch one by typing a slash command in the composer, browse the catalog from inside a project without dropping project scope, and hand new teammates starter prompts that already know which agent should run them.

Highlights:

• Skills in the main navigation: the skills catalog is a first-class workspace surface, no longer tucked behind Labs
• Slash-command launcher: type / in the composer to search and launch any skill inline, the fastest path from intent to running procedure
• Skills inside projects: a Skills tab in project context, so a project's relevant procedures are one click away
• Agent-aware starter prompts: prompt chips signal which agent they are built for, fall back gracefully, and can be bound per project
• Composer polish: icon-only pickers on mobile with a tint when a non-default option is active, and a setting to hide the default prompt chips for a clean slate

Product Intelligence: the dispatch knowledge connector moves in-platform

The product-intelligence connector for dispatch configuration, the deep knowledge base behind questions like "what does this setting actually do and what else does it touch", now runs on the platform's own infrastructure instead of an external host. The move brings the same locked-down access model as every other platform service, and it shipped with new capability: setting-interaction tools that trace how one configuration value affects others, with the evidence behind each claim.

Highlights:

• In-platform hosting: the knowledge service now runs on the platform's infrastructure behind service-to-service authentication, retiring the external endpoint
• Twelve tools: the full catalog covers settings lookup, search, comparison, and guidance for dispatch configuration work
• Setting interactions: ask how settings influence each other and get the interaction map plus the documentary evidence for each link
• Smarter routing: chat personas and corpus routing were refreshed across the full connector set, so questions land on the right knowledge source first
• Catalog visibility: the new tools appear in the workspace connector catalog with descriptions, so teams can see what is available before they ask

Platform: project lifecycle, proven restores, and safer database operations

This cycle's platform work focused on the unglamorous guarantees: what happens when data is deleted, and what happens when it must come back. Deleting a shared project is no longer a single destructive statement. It is a two-phase sweep that fences out concurrent writes, records every removed record class in the audit trail, and completes atomically or not at all. On the recovery side, a live restore drill proved that an on-demand database backup actually restores to a working instance, closing the loop on a promise most platforms only assume.

Highlights:

• Two-phase project deletion: an audited sweep with write fencing, so a delete cannot race against in-flight work or leave partial state behind
• Restore drill passed: an on-demand backup was restored end-to-end to a scratch instance, proving the recovery path rather than trusting it
• Bounded lock holds: the audit chain's integrity lock now runs under statement and idle timeouts, so a stuck writer cannot stall the platform
• Migration safety timeouts: every schema migration now runs with lock and statement timeouts, so a deploy cannot silently block production traffic
• Lifecycle design for what's next: archive, grace-window purge, restore, and ownership transfer are specified as the successor to hard delete

AI Studio: talk to your workspace

The workspace gained a voice, in both directions. Dictate a message instead of typing it, and let the assistant speak short progress updates back while longer work runs. Like the geo tools before it, the whole stack runs on the platform's own AI infrastructure: transcription and speech synthesis happen inside the perimeter, so voice input carrying sensitive operational detail is treated with the same care as anything typed.

Highlights:

• Composer dictation: tap the microphone, speak, and get clean text in the composer ready to review and send
• Push-to-talk: hold to record and release to transcribe, built for quick back-and-forth without touching the keyboard
• Spoken updates: during long agent runs, the assistant can speak brief status summaries so you can look away from the screen without losing the plot
• Pick a voice: choose from a curated set of voices in Settings, with playback speed control and a mute switch
• Perimeter-safe speech: transcription and text-to-speech are served by models inside the platform, not a third-party voice API

AI Studio: interactive maps and native geo tools

The workspace learned geography. Ask a question with a "where" in it and the assistant can now answer with an interactive map: pins for locations, drawn routes between stops, and shaded travel-time contours showing what is reachable within a set number of minutes. Behind the map sits a set of native geo tools (geocoding, routing, travel-time matrices, isochrones) served by the platform's own infrastructure, which matters because in this domain, addresses are sensitive data.

Highlights:

• Map artifacts: interactive maps rendered first-party in the workspace, with markers that carry contact details and website links, a matching table view, and pan and zoom like any map you know
• Routes and reach: draw the route between two or more stops, or shade the area reachable within 10, 20, or 30 minutes of a location
• Geo that stays home: geocoding, routing, matrices, and isochrones all run on services inside the platform's perimeter, so no third-party map API ever sees the addresses in your data
• Honest matching: when an address geocodes with low confidence, the result says so instead of silently guessing
• Keep and share: pin a map to a project like any artifact, and export it as a PNG image for use outside the workspace
• Fare-quoting APIs: address validation and point-to-point distance are also available as API endpoints for pricing and quoting integrations

Platform: adversarial security reviews and sharper redaction

Security work this cycle centered on adversarial review: multiple independent AI auditors, each from a different model family, were pointed at the platform's most sensitive code paths (tenant isolation, audit coverage, PHI leak paths, route authorization, file access) and told to break them. Their findings were cross-checked against each other, verified, and fixed, and the fixes shipped alongside a broader hardening sprint.

Highlights:

• Independent adversarial audits: several review rounds by different frontier models over the same hot paths, so a blind spot in one reviewer is caught by another
• Every confirmed finding fixed: verified findings were triaged into a hardening sprint and closed, with the fixes deployed
• Stricter redaction: the PHI scrubber now catches more identifier shapes and context patterns, including identifiers adjacent to names and dates that only matter next to a person
• Fail-closed connector boundaries: tools exposed by out-of-perimeter connectors are filtered fail-closed, so an unrecognized tool is excluded rather than allowed through
• Deeper isolation checks: forced row-level security extended further into the schema, with negative probes in the test suite that prove the app role cannot step outside its lane
• Safer deploys: database deploys now require a successful backup first, a reset guard blocks the one command shape that could wipe data, and a destructive-migration lint reviews every schema change

New connector: Resource Link community resources

The workspace can now search Resource Link, a curated directory of community organizations and the services they offer: food, housing, transportation assistance, health programs, and more, each with contact details and verification dates. Pin the connector to a chat or project and referrals in reports come from real, checkable listings instead of the model's memory.

Highlights:

• Thousands of verified listings: community organizations and their services, with addresses, contact information, websites, and the date each listing was last verified
• Read-only by design: the workspace can search and read the directory, never write to it
• Clean data boundary: the directory holds only public information about organizations and services, and it lives outside the platform's protected perimeter, so nothing about your members or clients ever reaches it
• In the catalog: Resource Link appears as a card in the Connectors catalog alongside the other curated connectors
• Works with maps: directory results carry locations and website links, so they plot naturally onto the workspace's new map artifacts

AI Studio: build your own agents and skills

The workspace's agent team is no longer fixed. Agents and Connectors moved into the main workspace navigation, and both can now be shaped to fit how your team works: customize the built-in agents, create new ones, and package repeatable procedures as skills any agent can use.

Highlights:

• Agent management: create, duplicate, and disable agents from the new Agents area; built-in agents can be customized, and disabling one hides it without losing anything
• Review before it ships: agent edits go through a draft-and-approve gate, so a change to an agent's instructions is always reviewed before it takes effect
• The right tools for each agent: every agent belongs to a capability class that bounds which models and tools it can use, so a custom agent cannot quietly gain access it should not have
• Guided skill authoring: a wizard walks you through writing a reusable skill, and a catalog shows what is available to your team
• Agents propose, owners approve: agents can suggest new project memories from a conversation, and the project owner approves or rejects each one before it is saved
• Connectors in view: each agent's connectors now show in the pickers, so it is clear what an agent can reach before you hand it work

AI Studio: task boards that run work

The project task board introduced earlier this month stopped being just a tracker. Cards can carry a prepared prompt, be assigned to an agent, and be run from the board itself, with the result landing back on the card. A Project Manager agent helps plan the work in the first place.

Highlights:

• Run a card: press Run and the assigned agent executes the card's prepared prompt on the server, with a durable record of each run and one run at a time per card
• Assign to people or agents: the assignee picker groups teammates and agents, so handing a task to an agent is the same motion as handing it to a colleague
• Charter to tasks: the Project Manager agent reads a project's charter and proposes a set of tasks for the board, which you review and accept
• Threaded comments: discuss a task on the card itself, with an authored comment thread on the detail view
• Better card details: a full-page detail view with editable fields, plus a one-click jump that opens the task in the assistant with the task's context already loaded
• Planner at hand: a floating planner on the board page keeps the planning conversation next to the work

AI Studio: attachments, saved prompts, and workspace organization

A wave of everyday-workflow upgrades landed across the workspace: getting material into a conversation, reusing what works, and keeping a growing workspace tidy.

Highlights:

• Chat attachments: drop files straight into a conversation, including .docx and .pptx documents, and the assistant reads them in place
• Saved prompts: a personal library of reusable prompts, with fillable fields for the parts that change each time
• Context meter: see how full a conversation's working context is, and when it runs long, summarize and hand off to a fresh context without losing the thread
• Chats and Artifacts pages: browse everything in one place, select ranges with shift-click, bulk delete chats, and bulk archive or export artifacts
• Artifact versions: revisions of the same artifact now group together with a version stepper instead of piling up as separate items
• Make it yours: a personal settings panel, your organization's logo in the app shell, and a dark theme by default

AI Studio: Projects gain tasks, memory, and files

Projects grew from a way to group chats into a real place to run work. Three new tabs sit alongside the chats: a board to track tasks, a memory that keeps a project's standing notes, and a files area that holds and previews the project's documents. A switcher in the sidebar moves between projects without leaving the flow.

Highlights:

• Task board: track a project's work as a simple board with columns from to-do through done, and move a card across as it progresses
• Project memory: keep durable notes and facts a project should carry between conversations, so context does not have to be re-explained each time
• Files tab: upload a single file or a whole folder, browse them as a tree, and preview Markdown, code, images, and PDF documents in the browser without downloading
• Sidebar project switcher: jump between your projects from the sidebar, with each project keeping its own chats, board, memory, and files
• Everything stays inside the workspace's isolated, encrypted storage, and file names and contents never appear in logs or the audit trail

AI Studio: image generation with provenance

AI Studio can now produce images, not just text and diagrams. Ask for an illustration or a visual and a specialist image agent creates it in the conversation. Because AI-generated media needs to be traceable, every image is stamped at the moment it is made.

Highlights:

• Generate visuals in a chat: describe what you need and a dedicated image agent produces it alongside the rest of the conversation
• Built-in provenance: each generated image carries embedded content credentials (the open C2PA standard) that mark it as AI-generated and record how it was made
• Honest metering: image generation is priced and metered per image and shown in the usage summary, so the cost of a visual is never hidden
• Picks the right tool: when a request calls for an image, the workspace routes it to the image agent automatically

AI Studio: governed web research

AI Studio can now reach out to the public web to ground an answer in current sources when you ask it to, and it cites what it used so you can check the work. Because a compliance-first platform cannot let a capability like this run unsupervised, it ships with an organization-level control and a clear record of when it was used.

Highlights:

• Grounded answers: when a question needs current information, a chat can search the public web and fold the findings into its reply, with each source cited inline
• Administrator control: web research is governed by an organization-level switch, so an administrator can enable or disable it for everyone at once
• A minimal, privacy-preserving record: the audit trail notes that a search happened and how many queries ran, never the query text or the results, so oversight never becomes its own data-collection risk
• Off by default where it should be: the capability stays dormant until it is explicitly turned on

AI Studio: take your work with you

Anything AI Studio produces can now leave the workspace in a format your team already uses. A generated report exports to PDF or Word, a table to Excel or CSV, and a set of slides to PowerPoint, so a chat result becomes a document you can share, edit, or file without copying and pasting.

Highlights:

• Office exports: save a generated document as PDF or Word, a table as Excel or CSV, and a slide deck as PowerPoint
• Diagram exports: export any generated diagram as a PNG image or an SVG vector file
• Full-screen diagram viewer: open a diagram to fill the screen, then zoom and pan to read the detail, and press Escape to close
• Every export is recorded in the audit trail the same way any other action on protected content is, so there is always a record of what left the workspace

Platform: self-serve onboarding groundwork

The groundwork for self-serve onboarding is in place. A fleet operator can create an account, choose products, accept the agreements, and land on a home base, with every order written to a permanent commercial record. The whole flow is built end to end and sits behind a coming-soon gate for now; it turns on once the customer database access is provisioned.

Highlights:

• Self-serve sign-up: create an account with name, work email, and organization name, using a password or Continue-with-Google, with the organization name turned into a clean, unique workspace identifier automatically
• Two separate agreements: the Terms of Service and the Business Associate Agreement are accepted and timestamped independently, never bundled into one checkbox
• Email-verified and fail-closed: nothing provisions until the email is confirmed, the payment is authorized, and both agreements are on record; the payment step is a clearly marked placeholder until the billing provider is connected, so the gate stays closed by default
• Permanent order records: a confirmed order is written to an append-only commercial stream (who ordered, which products and plan, the agreement versions and acceptance times, the payment method on file), kept entirely separate from the patient-data audit log and carrying no rider data
• Product-selection menu and home base: pick the products to enable and land on a home base showing your products, usage summary, notifications, and the module marketplace, reachable from a Home entry in the sidebar
• Usage feeds billing: AI usage is metered toward the billing record
• Gated for now: the sign-up and sign-in pages show a coming-soon note and the flow stays dormant, with no database calls, until access lands

AI Studio: Projects

AI Studio can now organize work into Projects. A project holds reference files that become standing context for every chat inside it, keeps its own artifact board, and can run a dedicated team of specialist agents, so each new conversation starts short and grounded instead of carrying a long history.

Highlights:

• Projects: group related chats under a project with its own list, detail view, and a start-a-chat-in-this-project action
• Reference files: add Markdown, plain-text, and CSV files one at a time or drop in a whole folder, subfolders included, in a single action; they become standing context for every chat in the project
• An honest pre-upload report: before anything uploads you see what will and will not be used, files skipped for type or size, and the ones that will not fit the project's context budget
• Per-project agent team: choose the specialists a project's chats delegate to, scoped to that project
• Artifact board: save a generated page, diagram, table, or document to the project and find it again on the project's Board tab
• File names and contents live in the workspace's isolated, encrypted storage and never appear in logs or the audit trail

A documentation site

Documentation moved onto its own footing. A dedicated documentation site now carries the product and onboarding content as its own service, with a clean knowledge-base layout, so the guides live somewhere stable instead of being scattered through the app.

Highlights:

• A documentation site served as its own service, with a clean knowledge-base layout for browsing and searching the guides
• Security and compliance posture documentation lives inside it as one reference, with old security links redirecting to the new home
• Consistent naming: product names were reconciled across the marketing site, the docs, and the changelog so the four-product story reads the same everywhere

AI Studio: agents, grounding, and multi-agent by default

AI Studio got clearer about who is answering and what they stand on. A new Agents page lists the specialist roster with the model each one runs, multi-agent delegation moved into the chat header and is on by default, and every grounded answer now shows its sources in plain sight.

Highlights:

• Agents roster: a page listing the specialist agents with a per-agent model picker limited to the approved Vertex AI models
• Multi-agent by default: the Subagents toggle moved into the chat header and now starts on, so a request that spans specialties fans out without hunting for a setting
• Sources in plain sight: when an answer draws on the regulatory corpus, the citations appear as a row of chips beneath it, each opening the full fact (statement, authority tier, jurisdiction, and source link) on hover
• Clearer grounding cards: the assistant's lookups now show numbered sources with inline citations instead of the earlier hard-to-read cards
• Knowledge-base connector: a read-only knowledge connector is wired into the assistant as an additional grounding source, alongside the regulatory corpus and the synthetic dispatch data
• The floating assistant now routes to a surface-aware agent, and its panel covers the whole screen on a phone

Platform: light theme, cross-surface navigation, and polish

Light mode got a ground-up redesign: neutral near-white surfaces with dark slate text, orange kept strictly as the accent color. The sidebar now groups pages by product line (Platform, AI Studio, Compliance Link, Business Intelligence) so the navigation matches how the product suite is described.

Highlights:

• Sidebar grouped by product line: Platform (audit log, usage), AI Studio (editor, chat), Compliance Link (compliance), Business Intelligence (fleet, dispatch)
• Portal home is now a launchpad of quick-link cards grouped the same way as the sidebar
• The portal, the landing site, and the docs link to each other; no more editing the address bar to switch surfaces
• Long conversations load fast: the chat loads the most recent messages first and offers a "load earlier" control
• Hydration warnings eliminated: the last source (chat date groups bucketing in server timezone) is fixed, so the console is clean on every hard load

Platform: audit integrity and tenant isolation proof

The audit trail now defends itself. Every audit entry is cryptographically chained to the one before it per tenant, so any after-the-fact edit, deletion, or reordering is detectable. An audit log browser page lets anyone with the right role read the trail with filters for action, resource type, actor, and date.

Highlights:

• Tamper-evident audit log: per-tenant SHA-256 hash chain; a verification tool walks the history and pinpoints the first broken entry
• Audit log browser: filters for action, resource type, actor, and date; entries show references (who did what to which record id), never values; viewing the log is itself an audited action
• Tenant switcher for testers: an allowlisted tester flips between the two synthetic tenants and watches chats, audit history, and usage change with zero bleed-through
• Tenant directory with proof of isolation: pages listing the workspace's tenants and showing each tenant's riders, drivers, and trips live from the isolated database, with encrypted fields shown as locked indicators
• Automated privacy guardrails in CI: every code change is scanned for PHI leaking into logs and for tenant context read from anywhere other than the verified sign-in

Platform: attestation PDFs and CI quality gates

Compliance paperwork is no longer manual. The platform fills the real Colorado attestation forms from trip data, producing signed, validated PDFs from a tolerant spreadsheet reader with a privacy-safe validation report (row numbers and counts, never values).

Highlights:

• Attestation PDF engine: fills Colorado mileage-attestation and long-distance verification forms from synthetic trip data; delivery into the portal UI comes next
• Quality gates now run automatically: a single check command bundles type, test, lint, privacy, and identity-seam gates; CI runs it on every push so the privacy guardrails no longer depend on someone remembering
• Behavioral test bench for the assistant: fixed conversations are replayed and scored on routing, delegation, synthesis, and resistance to prompt injection, including canary cases that prove rider details cannot leak into delegated task descriptions; runs with no cloud credentials
• Delegated work is billed at the specialist's own model rate, not the coordinator's, so cost totals on the usage page reflect what actually ran

Compliance Link: connector readiness matrix and role-scoped views

The connector catalog gained a readiness matrix: one view showing which tenant has which connector ready to enable, with live, coming-soon, and future integrations listed alongside the exact read-only tools each exposes. Write actions are excluded by design.

Highlights (Compliance Link):

• Connector readiness matrix: per-tenant enablement status across all connectors in the catalog
• Compliance browser confidence: the search result now shows which evidence tier each fact comes from, and topic searches no longer crash or return mislabeled results
• Jurisdiction filters work: federal and state filters were silently ignored before; they now narrow the results correctly

Highlights (Business Intelligence):

• Role-scoped operations views: the dispatch queue and fleet board now have a "view as" switcher (dispatcher, admin, driver) that limits the columns each role sees to what its job needs, the minimum-necessary access rule made visible on screen

Platform foundation on GCP

The first live version of the platform runs on Google Cloud. Cloud Run serves the portal, Cloud SQL Postgres holds the tenant data, and Identity-Aware Proxy gates every page to the tester allowlist.

Highlights:

• Multi-tenant Postgres with row-level security and field-level encryption on sensitive fields (signatures, notes, Medicaid IDs, dates of birth)
• Audit logging on every read and write of protected data, append-only from day one
• Vertex AI chat endpoint, streaming, with a Gemini model in the test environment
• Synthetic NEMT data so the platform can be built and demonstrated without any real patient data
• The entire test environment is reproducible from infrastructure code

AI Studio: full chat surface

AI Studio is now a full chat product. Conversations persist across visits, get automatic titles, and support rename, delete, share within the workspace, voting, and message editing. Answers render with headings, lists, tables, and syntax-highlighted code while they stream.

Highlights:

• Saved conversations grouped by date in a collapsible sidebar, with Ctrl-K search
• Model picker per message (server enforces the approved list), with Auto routing based on request weight
• Usage page: per-workspace totals and a day-by-day chart of turns, tokens, and estimated cost
• Grounded answers: the chat can consult the regulatory knowledge base with citations and look up synthetic dispatch data (trips, drivers), with live-state tools blocked
• The assistant can delegate: requests that span specialties fan out to specialist agents, with each delegated task visible in the conversation as it starts and finishes

AI Studio: editor and workbench

AI Studio now includes a full drafting editor with tables, code highlighting, slash commands, and inline AI assistance (ghost-text continuations and a selection menu for fix, extend, reduce, simplify, summarize, and translate). Documents save automatically and are isolated per tenant like everything else.

Highlights:

• Saved documents: create, rename, delete, and open documents with automatic saving and an editable title
• Images in the editor: drop or pick an image and it is stored in isolated, encrypted cloud storage and shown through an expiring link, with the upload recorded in the audit trail
• Workbench view: when the assistant fans a request to specialists, the conversation shows a live task list, a per-specialist report card, the sources it grounded on, and a per-turn cost readout
• Groundwork for chat file attachments: the platform stores and serves chat uploads in tenant-isolated, encrypted cloud storage with size, type, and content checks
• The assistant can now build things: ask for a page, diagram, table, or document and it appears as a card that opens into a side panel with a live preview

Compliance Link: compliance workspace and LEIE screening

The Compliance Link surface brings compliance work into one workspace. Search the tier-ranked regulatory corpus with jurisdiction and evidence-tier filters, open any fact's source document, and run OIG exclusion screening across the driver roster with a full paper trail.

Highlights:

• Regulatory corpus search with federal and state jurisdiction filters, evidence-tier badges (federal authority, state configuration, or context only), and abstention when no confident answer is available
• Monthly LEIE roster screening: one action screens the whole roster; exact registry matches are confirmed automatically while name-only matches go to a human review queue where each must be resolved with a written note
• Screening runs in resumable batches so an interrupted run can always be finished, with the Resume button shown only when a run has truly gone quiet
• Posture tab: how the platform classifies sensitive fields, BAA coverage status, and the current control posture
• A connector catalog with live, coming-soon, and future integrations, each disclosing exactly which read-only tools it exposes

Business Intelligence: fleet map, dispatch board, and analytics

Business Intelligence brings fleet and operations data into one product. A live fleet map shows status-colored vehicle pins paired with the asset roster. A dispatch board shows at-a-glance KPIs over live-feeling synthetic trips, alerts, and recurring schedules. Battlestation puts all of that on one display for the wall monitor and the demo stage.

Highlights:

• Fleet map: status-colored vehicle pins on a live map with region and status filters; selecting a row flies the map to that vehicle; operator metadata only, no rider data
• Dispatch board: fleet-on-board, active trips, SLA risk, open alerts, and on-time rate over synthetic fixtures with region and status filters that stay consistent across every panel
• Battlestation: a single-screen command board putting the KPI strip, trip queue, live fleet map, alerts feed, and roster on one display with region filters that cross-cut every panel
• Analytics suite: cross-system 360 view (including credentials expiring for on-shift drivers), operations deep-dive with annotated on-time-performance chart, operator leaderboard and payer mix, plus utilization and revenue views
• All charts run on deterministic synthetic data