Demo link updated: https://8b79-2804-1928-216d-4301-ed37-88a5-1969-d364.ngrok-free.app/

Included a pre-built Tester Agent that can be used in the public demo. This implements idea: DPI-I-439

Demo updated

The Memory card: your notes and your answers

• The status rail has a new Memory card, with two tabs: Notes and Your answers.
• Notes shows today and yesterday: one line for each of your messages, decisions, incidents, approval requests and reminders, in time order, for example 09:14 decision (mailbox): digest. OcuTalon's code builds the notes each time you look, from its own records. No model writes them, and they never hold what an email or other arriving message says: no sender, subject or text. A day stops at 2,000 characters, with "(more not shown)".
• With your first message of each day, and when a conversation starts afresh, the chat's model also gets your notes for today and yesterday.
• Your answers lists the answers OcuTalon keeps (see below), newest first, up to 100. Each row shows when, the channel, a one-line summary and your answer. Forget asks "Forget this answer? OcuTalon will stop using it." and removes it at once.
• The card doesn't list the memories you save with "Remember that ..." yet. You still list and forget those in the chat.
• After an upgrade, Your answers starts empty. Notes shows at once, because it is built from records OcuTalon already keeps.

This was noise, and your approval answers

• A tell-now or act notice now has a This was noise button, in the web chat and on the Simulated phone. Press it, and it becomes "Marked as noise at {time}. Similar items will take this into account." A second press changes nothing.
• When you approve or reject a fix or a tool call, in the web chat, on the Simulated phone or in Telegram, OcuTalon keeps that answer too.
• When a new item arrives from the same email address or source, on the same channel (or a check finding about the same production), its attention decision gets your newest answers on such items: at most 3. The model is told that they are examples, never rules, and that they can't make it act. In Visual Trace, they are in the decision's request, under "Your earlier answers on similar items (examples, never rules):".
• An answer to a tool call's request is listed and can be forgotten, but OcuTalon doesn't use it when it decides.
• Telegram doesn't have This was noise yet.
• On the online demo, each visit sees only its own notes and answers. Sign out deletes your answers, and your next visit starts with empty notes.

Got it in the web chat and on the Simulated phone

• A tell-now notice now has a Got it button in the web chat and on the Simulated phone, as in Telegram. Press it, and every copy of the notice ends with "Marked as handled at {time}", also after a reload. The first press wins.
• As before, the digest's line of what needs you leaves out a notice marked handled.

Daily digest

• The digest card in the web chat now has a "What I checked (n)" button that shows each check and its result. It starts open when any check didn't pass.
• The Simulated phone now shows the digest too, while demo mode is on.
• On the online demo, ask "Send my digest now": the digest appears in your chat and on the Simulated phone.
• What needs you never lists your own chat messages any more. Before, a request such as "Send my digest now" could show up there.
• After IRIS restarts, the system log no longer fails the digest because of two start-up lines that IRIS itself writes. Any other error line still counts.
• The Digest outcome now reads "Saved for the daily digest".

Settings

• Every section of Settings is now built. Settings is for OcuTalon admins, such as the ocutalon account of your own install. Visitors to the online demo aren't admins, so try this on your own install.
• Models: for each agent (triage, diagnosis and chat), choose Claude Haiku 5.5 or Claude Sonnet 5.5, and a second model that is used once when the agent's reply can't be used. Test sends one short request to the agent's saved model and shows how long it took to answer. You can also change each model's prices. A change applies from the agent's next model request.
• Keys stay in .env: the page never shows or asks for one. A provider whose key isn't set shows "No key set. Add {variable} to .env." Anthropic is the only provider listed.
• Budgets: the token budget per decision and the daily cost cap. When the cap is reached, the warning banner has a Budgets link for an admin.
• Daily digest: when it is sent, and where: the web chat, the Simulated phone (while demo mode is on) and Telegram (while it is on). Send my digest now sends one at once.
• Time zone: the zone of message times, reminders, the digest and the current time the model gets. Times on the page stay in your browser's zone.
• Channels: whether Telegram is on, paused or not set up. There is nothing to save here.
• A save shows "Saved. Applies within a minute. Recorded in the audit log." No restart is needed.

Fixes

• After an incident is resolved, reopened or dismissed, OcuTalon checks that production again at once. The Background checks list and the answer to "Is anything wrong right now?" agree with the incident within seconds, instead of up to a minute later. Try it on the online demo: approve the guide's fix and watch Background checks.
• An install or start could fail, saying an item didn't run, when that item was turned off by its default setting, such as the Telegram service on an install with no bot. The start's check now leaves such an item out.
• If IRIS stops while an approved call is running, OcuTalon now closes that call as interrupted when IRIS starts again, and never runs it a second time. Before, it could stay in progress for good.
• When an approved call can't run, the decision now says why in plain words: "No model key is set, so the approved call did not run. Nothing was changed." or "The model provider could not be reached, so the approved call did not run. Nothing was changed."
• The list of productions OcuTalon can watch now says whether each one is running, so the chat can tell you.

Data kept

• The nightly purge (03:00, 90 days by default) now also deletes chat turns and channel deliveries older than that, and a session with nothing newer. A session that is signed in, paired with Telegram, or has a watch, a pending reminder or an open approval request is kept, and so is one that holds a memory or a document. The purge never deletes memories, documents or IRIS audit log entries.
• Your answers and Got it marks older than 90 days are purged the same way.

Docs and install

• The README has a section for judges, with where to look for each criterion, a "New since 1.0.0" list with where to try each item on the online demo, and a contest checklist. Its Measured table now shows release 1.0.4's evaluation, and it says what changes without the agent: fixed rules alone get 6 of the 24 routine items right and give no incident a cause or a fix.
• The security overview says that daily notes and your earlier answers go to the model provider inside requests.
• The evaluation set has three new cases for your answers: the same email with no answers and with three This was noise answers from that sender, and an answer whose summary tries to give the model an instruction. They have their own score in the report, with no target yet.
• ./install.sh, the Docker image's build and the Python packages OcuTalon installs are unchanged since 1.0.4.

Requires the InterSystems IRIS for Health AI Hub Early Access image, build 154 or build 109. Upgrading from 1.0.4 keeps your conversations, reminders, watches, approval requests, settings, memories, documents and Telegram pairing. The online demo runs 1.0.5.

Initial Release

Frontend redesign for container and live demo. No Backend changes were added.

Shutter 1.3.2 is a patch on 1.3.0.

• The live banner on a phone: the budget line under the banner's bar stays on one line at 360 px.
• Tests: three test fixes, so the whole CI run passes again, including the online trial's deployment.
• README: the reason for building Shutter now says production exposes conditions that test environments cannot fully reproduce.
• Known issue carried from 1.3.0: the findings on an order or result case don't yet reach the case end to end on the test rig; the fix comes in 1.4.0, with pathology tier C.

Full change log: https://github.com/MxSalata/shutter/blob/main/CHANGELOG.md

Knos turns any research into a product billions of people use, in days, not years. Knos carries research across the valley of death, toward Kardashev II and III. Description updated.

Initial Release

Initial release of Allfedo's native IRIS AI Hub food-rescue workflow.

Includes signed-in kitchen, receiving, coordinator and collector views; reviewed food invitations; shared-capacity checks; actual handoff and receipt records; changed destinations; linked returns and disposal; and local SMTP notifications.

Tested on AI Hub Community 2027.1.0AI.109.0. Fresh installation, native checks and the complete browser journey passed. Demo partners and physical events are fictional.

Source release: https://github.com/snowsadh/iris/releases/tag/v0.1.0
Validation: https://github.com/snowsadh/iris/blob/main/doc/VALIDATION.md

README: For judges section; listing resources refreshed.

Included Demo functionality

Removed link for demo (under construction)

Shutter 1.3.0 adds pathology, tier B: when a receiver of orders and results has an outage, Shutter lists the messages that never reached it, for a person, and explains each failed order or result. Shutter still sends none of them.

• The resend set on an outage: each failed message of the window as queued, offered, reached or refused, with its reason and a check of every patient on it; recomputed each minute and on Read again.
• What Shutter found on an order or result case: the receiver's own words field by field, the OBX they name, and where a character above plain ASCII stands. The escalation draft names only the segment, field and code, never a result's value.
• The outage email, opt-in per receiver, naming the receiver and the counts, never a patient.
• Outages on the Incidents page, above the incidents.
• A fix: a message read in a hospital's namespace is copied whole, so the patient check always reads the right message.
• FHIR R4 is in development: its plan builder, verifier and a test FHIR server are in the release and tested; no case uses them yet.
• Also: the evaluation's price table carries Sonnet's current cache-read price.

Full change log: https://github.com/MxSalata/shutter/blob/main/CHANGELOG.md

Shutter 1.3.0 adds pathology, tier B: when a receiver of orders and results has an outage, Shutter lists the messages that never reached it, for a person, and explains each failed order or result. Shutter still sends none of them.

• The resend set on an outage: each failed message of the window as queued, offered, reached or refused, with its reason and a check of every patient on it; recomputed each minute and on Read again.
• What Shutter found on an order or result case: the receiver's own words field by field, the OBX they name, and where a character above plain ASCII stands. The escalation draft names only the segment, field and code, never a result's value.
• The outage email, opt-in per receiver, naming the receiver and the counts, never a patient.
• Outages on the Incidents page, above the incidents.
• A fix: a message read in a hospital's namespace is copied whole, so the patient check always reads the right message.
• FHIR R4 is in development: its plan builder, verifier and a test FHIR server are in the release and tested; no case uses them yet.
• Also: the evaluation's price table carries Sonnet's current cache-read price.

Full change log: https://github.com/MxSalata/shutter/blob/main/CHANGELOG.md

Shutter 1.2.1 fixes what the checks run on 1.2.0 found on IRIS 2026.2 and AI Hub build 154.

• Two people deciding the same approval at once: the second is told the decision that stands, with who took it and when.
• A cancel waiting for an approval's lock no longer leaves a runtime error in the Event Log.
• Six tables keep their storage map in their sources, so a later upgrade reads every existing row as before.
• The cases list is read once per visit, and again after a decision.
• The README: clearer steps for judges, Ask Shutter, the API page and document support (MDM^T01 to T11 and report-carrying ORU^R01) named, and the second and third articles linked.

Full change log: https://github.com/MxSalata/shutter/blob/main/CHANGELOG.md

Shutter 1.2.0 brings everything since 1.0.1 to Open Exchange: an assistant that answers questions about the screen you're on, the tools a hospital needs before a pilot, outage watching, clearer screens, and fixes from three rounds of independent review. The safety rule hasn't moved: the AI board proposes, a deterministic gate decides, and a person or the site's written policy releases every write.

New for you to try

• Ask Shutter: a panel beside every page that answers questions about what's on screen. On a case, About this case opens it with the case's own questions; it is never recorded or replayed.
• Clearer screens: the case page shows the key facts first, with click-to-define terms. The Technical view adds the HL7 message type, the receiver, the source system and message id, and the HTTP status. Each disposition has its own colour and icon.
• API page: every route in the Technical view, backed by a published OpenAPI 3.1 description that CI checks against the code.
• Also new: the Health page separates its checks from information, Settings compares what your account can do with the other roles, and the live hospital banner closes for the session and comes back when an incident opens.

For a hospital getting ready to pilot it

• Shadow mode and Stop Shutter recovery: a site install, and an upgrade whose profile doesn't say otherwise, starts in shadow mode: Shutter investigates and records what it would do, and writes nothing until the site enables recovery. One switch holds every recovery call until an approver clears it.
• Receiver checks: a conformance kit for the receiver's supplier, a list of what to check at the receiver on every stopped case, and warehouse history for a receiver with no API.
• The rest of the toolkit: a failure coverage census, document exceptions (MDM and report-carrying ORU), a pilot pack (decision record, stage plan, hazard log, data flow) and a case bundle for audit.
• Outage watching (1.1.0): Shutter watches HL7 TCP operations without contacting the receiver, groups an outage into one incident, and puts an incident of any size under one approval.

Safer recovery

• Shutter stops for a person wherever a resend could be wrong: a later movement on another visit, a death on the message, two systems moving a patient at once, a receiver with no readable history.
• An upgrade keeps a site's stop for another visit's movement; only a new install shows it on the plan instead.
• A null bed in a receipt no longer holds a settled case.
• A supported uninstall takes off everything the install added, and refuses rather than harm a site's own production.

Tested

• Three independent code reviews, every finding fixed before the release.
• Unit suites and the board scenarios on IRIS 2026.2 and AI Hub build 154; install and unit suite on build 109.
• The judges' replay install checked end to end on 2026.2.

Full change log: https://github.com/MxSalata/shutter/blob/main/CHANGELOG.md

DICOM3D AI Hub — Initial Public Release (v0.1.0)

DICOM3D AI Hub is a research imaging application demonstrating native InterSystems AI Hub agents within an IRIS for Health Interoperability workflow.

Features

• Native AI Hub agent integration using %AI.Agent.
• Retrieval of restricted FHIR R5 imaging evidence.
• Evidence validation and source-linked anatomical structure selection.
• Synthetic CT/SEG mismatch detection with a review-request outcome.
• Interactive 3D anatomy and medical imaging exploration.
• Expanded male reference anatomy.
• Scientific, Holographic, and Materialization viewing modes.
• Standalone Safe Demo source with clearly labeled simulated AI outcomes.

Installation and Reproducibility

Includes source code, installation instructions, research-data attribution, and documented qualification evidence.

InterSystems Community AI Hub Build 154, required FHIR images, and model weights are obtained separately and are not redistributed.

Current Limitations

This is a research prototype, not a clinical diagnostic application.

The Safe Demo is not yet hosted and currently requires WebGPU. A WebGL fallback remains unqualified.

Native AI Hub execution has been demonstrated, but intermittent inference failures and incomplete repeatability testing remain disclosed.

Source: https://github.com/KennethJSmithDev/DICOM3D_AI

Demo link

Documents

• Add your own documents on the Documents page: press Add documents and pick Markdown, plain-text, PDF or Word (.docx) files. Each one shows "Indexing…" and then "Ready".
• Then ask about them in the chat, for example "How do I rotate the Results certificate?" OcuTalon searches your documents, answers from the passages it finds, and names the document and section it used. A Markdown file's sections are its ## headings, a PDF's are its pages (Page 1, Page 2 and so on) and a Word file's are its Heading 1 headings.
• When no passage answers the question, it replies "None of your documents cover that."
• OcuTalon's code, not the model, records which passages an answer used: at most 5, and only passages that match the question closely enough. When an answer has any, OcuTalon takes "None of your documents cover that." out of it, so the two never appear together. The web chat doesn't show these citations under the answer yet.
• A file that can't be read stays in the list as "Couldn't read it:" and the reason: "it isn't a Markdown, plain-text, PDF or Word (.docx) file.", "that file is {size}; the limit is 10 MB.", "the file is empty.", "OcuTalon found no text in it.", "it's protected with a password.", "it's damaged or incomplete.", "it has no text to read, for example a scanned PDF without a text layer.", "it's an old Word file (.doc); save it as .docx and try again.", "its text is too long to read.", "reading it took too long." or "something went wrong inside OcuTalon."
• The limits: 10 MB for a file of any type. OcuTalon reads the first 200,000 characters of a PDF's or a Word file's text, and refuses one whose contents unpack to more than 100 MB ("its text is too long to read."). A Markdown or plain-text file is read whole, up to IRIS's longest string (about 3.5 MB); a larger one shows "something went wrong inside OcuTalon." PDF and Word files are read one at a time, and one that isn't read within 45 seconds, waiting included, shows "reading it took too long."
• What a document says is data, never instructions: once OcuTalon has searched your documents, the rest of that turn runs only tools that read, so in that turn it can't, for example, save a memory or set a reminder.
• The passages become vectors on your instance with the same FastEmbed model as memories (no key, and no network once installed), and are searched with AI Hub's knowledge base over an IRIS vector store.
• On your own install, documents belong to the install: every user of the instance sees and searches the same ones.
• On the online demo, each visit has its own documents and memories: no other visitor can list or search them. Sign out deletes them; otherwise they are kept until the demo's next reset. Visual Trace on the demo is shared, so don't add anything personal there.

Daily digest

• The digest now shows in the web chat, on every install, as a card with the same lines as the Telegram digest. It is still there after a reload.
• On your own install, the 08:00 digest goes to the web chat, and to Telegram too when it is paired.
• On the online demo, ask "Send my digest now" and the digest appears in your chat. The demo sends no morning digest: a visitor gets one only by asking.
• The Simulated phone doesn't show the digest, and the Daily digest section of Settings isn't built yet.

Got it on Telegram

• On your own install with Telegram paired, the Got it button under a tell-now message now works. Press it, and OcuTalon marks that notice handled, ends the message with "Marked as handled at {time}" and removes the button. A second press changes nothing.
• The digest's line of what needs you leaves out a notice marked handled. It counts tell-now and act decisions only from the time of the previous digest (the last 24 hours, for the first one); failed checks and incidents waiting for your approval always count.

Fixes

• Remember and forget: OcuTalon's attention decision on a "Remember that ..." message could come back without a confidence, so it fell back and showed the Fallback flag. The instructions now ask for a confidence on a request to remember or forget something, as for any other message.
• Several incidents at once: two diagnoses now run at the same time instead of one, so when incidents arrive together, fewer of them spend their timer waiting in a queue. A request whose decision is already recorded, for example after its timer ran out, is skipped when its turn comes: no model call and no cost.

Demo guide

• The "Other things to try" panel now lists Ask about a document. Its button, Open the Documents page, takes you there.

Docs and install

• The security overview lists the new documents tool and says how it is classified: safe and read-only, but its output is never trusted, and each search returns at most 5 passages.
• An install or upgrade adds the documents entry to the ToolClassification setting of Decide and keeps every other entry there.
• The bridge guide describes a fourth extension class, TurnToolsClass, and the reply filter's new form, Filter(text, hostData).
• ./install.sh, the Docker image's build and the Python packages OcuTalon installs are unchanged since 1.0.3.

Requires the InterSystems IRIS for Health AI Hub Early Access image, build 154 or build 109. Upgrading from 1.0.3 keeps your conversations, reminders, watches, approval requests, settings, memories and Telegram pairing. The online demo runs 1.0.4.

Initial Release

Initial Release

Initial Release

Fixed some IPM install issues when trying on different namespaces.

Initial Release

Initial release for the InterSystems "AI Agents" contest.

• IRIS AI Hub agent reads each new FDA drug recall inside an Interoperability production and decides: patient contact, shelf check, or no action.
• Finds which patients received the recalled drug and ranks them by estimated chance (exact when a scanned lot matches).
• Pharmacist approval through the IRIS Workflow engine: expires in 10 minutes, single use; kill switch; full audit trail with Visual Trace.
• Live accuracy on 300 real FDA recalls with claude-haiku-5-5: ingredient 98.3% (rule-based 82.0%), 0 invented NDCs, total cost about $0.28.
• Runs with docker compose up; without an API key it replays the recorded Claude outputs.

This update adds a full InterSystems IRIS Interoperability production, Copilot.Interop.Production. In it, the AI agents powered by InterSystems AI Hub are regular workflow components, with queues, message trace, audit and auto-start.

Ask workflow: every question asked in the portal

Copilot Ask Service → Copilot Ask Process → AI SQL Agent ⇄ IRIS SQL Executor → Audit Log

The AI SQL Agent (business operation, %AI.Provider) turns the question into IRIS SQL, using the database catalog and the business glossary.
The IRIS SQL Executor validates the query (read-only guard) and runs it.
The self-correction loop is explicit. When IRIS returns an error or an empty result, the business process sends the real IRIS error back to the agent as a new message, up to 3 attempts. Every attempt is visible in the message trace.
Every interaction is recorded by the Audit Log operation.
Report and alert workflow: scheduled or on demand (BPL)

Report Scheduler (Ens.InboundAdapter) / Report Trigger → Report Process (BPL) → IRIS SQL Executor → alert rule → AI Report Agent → Report Publisher → Webhook Notifier → Audit Log

Users schedule any result as a report, with a frequency, optional alert thresholds and an optional webhook.
The AI Report Agent runs the autonomous %AI.Agent (with its %AI.ToolSet) to write the report narrative.
The Report Publisher writes each report to disk through EnsLib.File.OutboundAdapter and shows it in the portal. The data table always comes from IRIS, never from the LLM.
The Webhook Notifier posts alerts and reports to external systems such as Microsoft Teams, Slack or automation tools.
A sample “Daily revenue check” subscription runs at startup, so a report and its trace are available on first launch.
Visibility
Workflows page in the portal: production status, live message counts per component, scheduled reports with Run now, generated reports and a built-in message trace viewer.
Every answer shows a 🔀 session N badge that opens the trace of that question.
IRIS Management Portal: the AI Hub images ship without a web server, so a new InterSystems Web Gateway container serves the standard portal at http://localhost:52773/csp/sys/UtilHome.csp. It gives access to the production configuration, the BPL editor, the Message Viewer and the Visual Trace.
Security and privacy
User privileges preserved: production jobs run under the IRIS service account, so the catalog is built in the signed-in user’s own process and the validated SQL is executed again as that user before any data is shown.
Privacy mode (PRIVACY_MODE=1): no data value is ever sent to the LLM, only the question and metadata. Result analyses use aggregated statistics only, and none for results smaller than 10 rows, to prevent re-identification. This supports LGPD/GDPR-sensitive scenarios.
Graceful fallback: if the production is stopped, the portal keeps working through direct calls.
Validation

Rebuilt from scratch and tested end to end on iris:2027.1.0AI.108.0:

Interoperability is enabled in the namespace and the production and BPL compile without errors.
The production auto-starts.
Both workflows completed successfully, including the self-correction loop and a triggered alert.
The Management Portal Visual Trace and production configuration were checked in the browser.

Repository: https://github.com/Andre-Dienes/iris-dashboard-ai (see the “AI agents in interoperability workflows” section of the README).

Initial Release

Initial release: IRIS AI Hub agent proposals, independent coding checks, guided algebra practice, and Docker installation verified.Initial Release

IPM package published to the public registry: zpm "install interop-ops-agent". module.xml version 0.2.2, no code change.