Updated model access

Updated models

Updated agents integration with mimo model from xiaomi

Initial Release

wire up real LLM inference engine

• agents now powered by Cloudflare Workers AI

Fixes Applied per Review Feedback

📝 Documentation Updates & Release Details (v2.1.1)

We have updated the project's README.md and key pages of the claimauditai-wiki to document production-grade hardening, security fallbacks, and the newly integrated Model Context Protocol (MCP) server.

---

🚀 1. Model Context Protocol (MCP) Terminology Server

• Documentation: Terminology MCP Server Wiki & README.md Details
• Implementation: mcp_server.py
• Details: Documents the new FastMCP terminology resolver deployed inside the claimaudit-iris container. It enables the AI chat assistant to call standardized local clinical tools to prevent LLM hallucinations:
  • lookup_cpt_code(code: str) -> str: Translates 5-digit CPT codes to official procedure descriptions.
  • lookup_icd_code(code: str) -> str: Translates diagnosis codes, supporting prefix matches.
  • validate_diagnosis_procedure(icd_code: str, cpt_code: str) -> str: Runs clinical compatibility audits.

---

🤖 2. Machine Learning Drift Protection & Safeguards

• Documentation: Autoencoder Architecture Wiki & README.md ML section
• Implementation: autoencoder_train.py
• Details: Documents model training protections introduced to ensure stability in production environments:
  • Validation Split ($15%$): During retraining, $15%$ of historical data is reserved for evaluation, and training is executed on the remaining $85%$.
  • Drift Threshold ($15%$): The candidate model's validation loss must not degrade by more than $15%$ compared to the active baseline model, or the update is safely rejected.
  • Rolling Backups: Automatically rotates and retains the last 3 successful model weights and stats (_v1.pth/npz, _v2.pth/npz, _v3.pth/npz) for instant recovery.
  • Weight Determinism: Details the usage of torch.manual_seed(42) before model instantiation to eliminate non-deterministic sequential loss fluctuations.
  • Bypass Safety: Gracefully bypasses evaluation (returning flagged=False) when there are fewer than 5 historical claims instead of throwing errors.

---

🔑 3. Non-Blocking JWT Security Fallback (Safe Mode)

• Documentation: API Key Handling Wiki & README.md Security section
• Implementation: Auth.cls
• Details: To guarantee reviewers can inspect and access the system without configuration blocks, the application now boots in Non-Blocking JWT Safe Mode if JWT_SECRET is missing.
  • If missing, the system logs a warning to the ^ClaimAuditSecurityError global and falls back to a persistent GUID rather than causing fatal startup exceptions.

---

📊 4. Direct Stats Extraction & Transaction Queues

• Documentation: README.md Adjudication Queue section
• Implementation: Router.cls & Queue.cls
• Details:
  • Extension Parsing: Refactored the dashboard statistics logic to directly parse structured https://claimauditai.com/fhir/extension/tier-results JSON objects inside ClaimResponse instead of using brittle text substrings.
  • Background Processing: Documents the asynchronous queue (ClaimAudit.Data.Queue) processing engine, which allows immediate 202 Accepted ingestion responses for clients while executing AI pipelines in background worker threads.

---

🧪 5. Expanded Verification & Testing Suites

• Documentation: README.md Testing section
• Implementation: test_mcp_server.py
• Details:
  • Python unit test cases have been expanded to 92 tests, verifying the MCP endpoints, drift validations, backup rotations, and graph cycles.
  • All Vitest, Pytest, and end-to-end integration tests have been validated to run successfully.

---

📚 New Wiki Documentation & Accuracy Audit

Recent commits (b0ffe6d, 081bf4f, and feda00f) added 6 new architectural documents to the claimauditai-wiki and updated the README.md to establish complete alignment between the codebase and the documentation:

• Data Classes Wiki — Documents the persistent storage classes (ChatHistory with reserved keyword escaping, background Queue, Cytoscape-compatible GraphStore, and Debug utilities).
• UI Architecture Wiki — Details the frontend React 19 + TypeScript + Vite 6 tech stack, Zustand 5 state stores, and TanStack Query 5 data-fetching structures.
• LLM Router Architecture Wiki — Explains the failover queues, exponential backoff retries, sliding window rate limits, caching TTL, and SSE chat streaming in the backend.
• Agent Tool Registry Wiki — Covers the decorator-based function introspection schema generation for tools mapped to the ReAct agent loop.
• Diagnosis-Procedure Validator Wiki — Details rule-based CPT and ICD-10 compatibility audits and validation chapter range mappings.
• FHIR Extension Definitions Wiki — Documents custom billing, decision, and risk scoring extensions attached to pended ClaimResponse resources.

ClaimAuditAI — Release Notes v2.1.0 (Production Hardening & Drift Safeguards)

Welcome to the v2.1.0 Release of ClaimAuditAI. This release focuses on production-grade hardening, model drift protection, training determinism, and robust integrations with the Model Context Protocol (MCP) terminology server and structured FHIR extensions.

---

🛠 Architectural & Safety Enhancements

1. 🤖 PyTorch Autoencoder Drift Checking & Validation Split

To prevent model degradation in production, we introduced a robust verification loop during the autoencoder retraining process:

• Validation Split (15%): Retraining now reserves 15% of historical claims as a validation dataset, with the remaining 85% used for training.
• Validation Drift Safeguards (15% Threshold): Prior to overwriting the active model weights (autoencoder_model.pth), the candidate model is evaluated on the validation dataset. If the candidate loss represents a degradation of more than 15% compared to the baseline model's loss, the update is rejected, a warning is logged, and the previous model remains active.
• Rolling Backups (3 Versions): The system automatically rotates and maintains the last 3 successful model weights and stats (_v1.pth/npz, _v2.pth/npz, _v3.pth/npz) for instant rollback capability.
• Deterministic Weight Initialization: Resetting the seed (torch.manual_seed(42)) immediately before candidate model instantiation guarantees deterministic initialization, preventing false drift rejections in automated sequential test runners.

2. 🔑 Non-Blocking JWT Secret Review Fallback

• Location: [Auth.cls](file:///Users/mck/Desktop/claimauditai/src/cls/ClaimAudit/REST/Auth.cls)
• Hardening: Modified GetJWTSecret() to not block application startup or throw a fatal exception if CLAIMAUDIT_ENV=production is set but the JWT_SECRET environment variable is missing.
• Auditing: Logs a clear warning to the ^ClaimAuditSecurityError global and falls back to a persistent GUID. This ensures the application remains functional for contest reviewers while keeping a clear audit trail of security configuration gaps.

3. 📊 Direct Stats Parsing via FHIR Extensions

• Location: [Router.cls](file:///Users/mck/Desktop/claimauditai/src/cls/ClaimAudit/REST/Router.cls)
• Hardening: Refactored GetStats() to extract the structured https://claimauditai.com/fhir/extension/tier-results extension directly from the pended ClaimResponse resource.
• Accuracy: Replaced brittle substring parsing of the human-readable disposition text with deterministic JSON parsing of the flagged status for tier1, tier2, and tier3. This ensures the dashboard stats donut chart accurately counts flags regardless of formatting adjustments.

💬 Conversational Terminology Lookup

• Exposed FastMCP tools (lookup_cpt_code, lookup_icd_code, validate_clinical_edits) to the Chat Agent loop. When a user asks about a CPT/ICD code description in the UI Chat Assistant, the LLM issues a tool call to query the local terminology server, returning official definitions instead of generating hallucinations.

---

🚦 Verification Status

• Python Unit Tests: All 92 tests passed successfully, including the validation split, drift check, and backup rotation tests:

  $ pytest src/python/tests/
  ======================== 92 passed in 85.85s =========================
  

• Vite/React Unit Tests: Vitest suite completes successfully with 18 passed tests (100% success rate).
• End-to-End Workflow: Real-world integration tests successfully verify data clearing, seeding, retraining, audit hold escalation, director approvals, and localized graph analysis.

ClaimAuditAI — Release Notes v2.0.0 (Agentic Payment Integrity)

Welcome to the v2.0.0 Release of ClaimAuditAI. This release marks a major shift, transitioning the payment integrity engine from a sequential ReAct tool-calling loop into a type-safe, compiled Agentic Finite State Machine (FSM) powered by Pydantic Graph.

Additionally, this version incorporates advanced InterSystems IRIS for Health integrations, custom Model Context Protocol (MCP) terminology engines, and Explainable AI (XAI) evidence citation linkages.

---

🛠 Major Architectural Upgrades

1. Pydantic Graph FSM & Agentic Memory

We refactored the Python-based adjudication pipeline into a type-safe, compiled FSM (agent_graph.py):

• State-driven Adjudication (AuditState): Explicitly type-checked execution context tracking input metadata, tier-specific findings, citations, and LLM synthesis results.
• Structured Nodes: Implemented as independent classes inheriting from BaseNode[AuditState, None, str] with StepContext validation:
  • ClaimIngestionNode: Ingests and sanitizes metadata.
  • ClinicalAuditNode: Executes vector database clinical note alignment.
  • AnomalyAuditNode: Calculates autoencoder reconstruction loss.
  • NetworkAuditNode: Evaluates referral loops and address collisions on the provider graph.
  • LLMSynthesisNode: Aggregates scores and builds the markdown report.
• Deterministic Fast-Approval Bypass: Claims returning clean metrics across all three tiers bypass the LLM synthesis node entirely, transitioning directly to End to save API tokens and reduce latency.

2. Hybrid Agent Bridge Wrapper

• Class: ClaimAudit.AI.AgentWrapper
• Integration: Dynamically checks for the presence of native InterSystems %AI.Agent classes via class dictionary compilation. If present, it maps tools as %AI.Tool bindings under %AI.ToolSet to run natively within AI Hub. Otherwise, it falls back to execution via the Python-compiled Pydantic Graph FSM.

3. Unified Python Tool Registry

• File: [agent_tools.py](file:///Users/mck/Desktop/claimauditai/src/python/agent_tools.py)
• Implementation: Created a decorator-based tool registry (@registry.register) that dynamically generates OpenAPI/JSON schemas using function inspection. Exposes 8 medical and diagnostic tools:
  • lookup_cpt_code / lookup_icd_code (resolves medical codes).
  • validate_clinical_edits (verifies diagnosis-procedure combinations).
  • run_nlp_audit / run_anomaly_audit / run_graph_audit (live tier executions).
  • get_patient_history / get_provider_history (historical database lookups).

---

💡 Contest & InterSystems Enhancements

1. Terminology Model Context Protocol (MCP) Server

• Implementation: Developed [mcp_server.py](file:///Users/mck/Desktop/claimauditai/src/python/mcp_server.py) utilizing FastMCP to expose terminology resolution services.
• Functionality: Allows LLM models and the UI chat assistant to resolve medical codes dynamically and validate clinical relationships (CPT/ICD mapping).

2. FHIR SQL Builder Integration

• Projections: Schema definition file [fhirsql/projections.json](file:///Users/mck/Desktop/claimauditai/fhirsql/projections.json) enables standard projected tables for Claim and ClaimResponse resources.
• Helper Class: Compiled [SQLBuilderHelper.cls](file:///Users/mck/Desktop/claimauditai/src/cls/ClaimAudit/FHIR/SQLBuilderHelper.cls) to execute dynamic SQL statements over projections, eliminating manual JSON parsing queries.

3. Explainable AI (XAI) Evidence Citations

• Evidence Capturing: Modified the Python ML tiers to extract and bubble up absolute resource ID references (e.g. DocumentReference ID for clinical notes, Practitioner NPIs for address collisions).
• Database Mapping: Added the ClaimId column to the ClaimProjections database table, populated it during ingestion, and stored array citations in the FHIR tier-results extension on ClaimResponse.
• UI Presentation: Displays evidence citations as styled pill tags with hover effects underneath the tier results on the claim detail page.

---

🔒 Production Hardening & Bug Fixes

• Concurrency Safety: Replaced multi-threaded execution in tier_orchestrator.py with sequential processing to eliminate IRIS Embedded Python database context conflicts.
• Autoencoder Scaling: Scaled SpecialtyCode to [0,1] prior to evaluation, preventing categorical outliers from skewing continuous Z-score distances.
• JIT LLM Adjudication: Configured bulk seeding to bypass LLM generation, compiling detailed markdown reports on-demand (JIT) when an auditor actually views the claim detail.
• Ledger Linkage: Added claimResponseId to ledger records, converting Claim IDs in the Override Ledger into navigation links back to resolved read-only detail views.

---

🚦 Verification Status

• Python Unit Tests: All 90 pytest tests pass cleanly, including the new Pydantic Graph node transition and mock fallback suites.
• Vite/React Unit Tests: Vitest suite completes successfully with coverage reporting enabled.
• TypeScript & Lints: Clean compilations (tsc --noEmit exits with 0).
• End-to-End Workflow: Real-world integration tests successfully verify data clearing, seeding, retraining, audit escalation, and director decisions.

Here is the release description for ClaimAuditAI v1.0.2, formatted for immediate use in GitHub Releases, community announcements, or project documentation:

---

🚀 ClaimAuditAI v1.0.2-Enhancements Release

This release introduces three key architectural enhancements: FHIR SQL Builder Integration, a dedicated Medical Terminology MCP Server, and Explainable AI (XAI) Citations to link adjudication findings directly back to their source clinical and administrative evidence.

---

🌟 Key Highlights

1. ⚙️ FHIR SQL Builder Integration

Allows developers and administrators to bypass manual database projections in favor of platform-native relational maps for advanced analytics.

• fhirsql/projections.json: Pre-configured JSON projection configuration for Claim and ClaimResponse resources. This file can be imported directly into the InterSystems FHIR SQL Builder Management Portal.
• ClaimAudit.FHIR.SQLBuilderHelper: A custom ObjectScript query helper that executes dynamic SQL queries against projected schemas, extracting CPT codes, ICD diagnoses, amounts, and dates with built-in schema fallbacks.

🤖 2. Medical Terminology FastMCP Server

A Model Context Protocol (MCP) server built with Python's FastMCP framework, enabling cognitive agents to perform standard clinical translations and diagnosis-procedure justification checks.

• lookup_cpt_code(code): Resolves a 5-digit CPT code into its human-readable clinical procedure description.
• lookup_icd_code(code): Translates an ICD-10 diagnosis code into its standard clinical description.
• validate_codes(icd_code, cpt_code): Performs programmatic validation to ensure the diagnosis supports the billed procedure.

🔍 3. Explainable AI (XAI) Citations

Links threat signals back to specific, traceable FHIR resource identifiers in the adjudication trail, giving auditors exact references to support hold decisions.

• Tier 1 (NLP): Captures the DocumentReference/{id} identifier of the progress note matching the billed CPT code.
• Tier 2 (ML): Citations link directly back to the audited Claim/{id} as evidence of statistical outlier loss.
• Tier 3 (Graph): Returns conflicting Practitioner/{npi} and Claim/{id} references for geodetic/temporal leaps or address collisions.
• UI Integration: Displays evidence citations as clean, monospace pill tags featuring color-coded indicator dots matching the active tier and subtle scale transitions on hover.

---

🛠️ Technical Changelog

🗄️ Database & ObjectScript Layer

• Updated [Engine.cls](file:///Users/mck/Desktop/claimauditai/src/cls/ClaimAudit/AI/Engine.cls) table schema definition to add ClaimId to the ClaimProjections table with a safe ALTER TABLE schema update fallback.
• Updated [Router.cls](file:///Users/mck/Desktop/claimauditai/src/cls/ClaimAudit/REST/Router.cls) and [Engine.cls](file:///Users/mck/Desktop/claimauditai/src/cls/ClaimAudit/AI/Engine.cls) to parse python citations lists and persist them into the dynamic tier-results extension array.

🐍 Python Adjudication Pipelines

• Added mcp>=1.27.2 to [requirements.txt](file:///Users/mck/Desktop/claimauditai/requirements.txt).
• Updated [nlp_auditor.py](file:///Users/mck/Desktop/claimauditai/src/python/nlp_auditor.py) to select and return note references DocumentReferenceId in native vector search.
• Updated [graph_analyzer.py](file:///Users/mck/Desktop/claimauditai/src/python/graph_analyzer.py) to read ClaimId from projected edges and return conflicting NPIs and claims for address collisions.
• Updated [tier_orchestrator.py](file:///Users/mck/Desktop/claimauditai/src/python/tier_orchestrator.py) to pass claim_id parameters and forward citations downstream.

💻 React UI Dashboard

• Added citations type interfaces in [claim.ts](file:///Users/mck/Desktop/claimauditai/ui/src/types/claim.ts).
• Styled and integrated citations rendering in [TierPanel.tsx](file:///Users/mck/Desktop/claimauditai/ui/src/components/claims/TierPanel.tsx) with responsive padding and micro-interactions.

📚 Documentation & Wiki

• Added the [Terminology MCP Server](file:///Users/mck/Desktop/claimauditai/claimauditai-wiki/07-api/Terminology%20MCP%20Server.md) wiki page.
• Updated the [FHIR SQL Builder Projections](file:///Users/mck/Desktop/claimauditai/claimauditai-wiki/04-fhir/FHIR%20SQL%20Builder%20Projections.md) wiki page and the MOC home index.

---

🚦 Verification Status

• Python Tests: 82 passed unit tests (including the new test_mcp_server.py suite).
• UI Code Quality: TypeScript check compiled cleanly with zero errors (npx tsc --noEmit exited with 0).
• Workflow integration: Successfully executed real_world_e2e_tests.py against the running docker containers to verify database seeding, autoencoder training, single-claim details fetching, and correct REST response mapping.

• missing module.xml update ...

• Fix: Replace GetHeader("Content-Type") by using the Property Content-Type (%Net.HttpRequest). GetHeader("Content-Type") does not work because it's not forwarded like the SetHeader method.

• added a frontend for testing and viewing results
• added project live

Readme updates and minor fixes

Release Notes: ClaimAuditAI v1.0.1

This release focuses on production-hardening the containerized deployment lifecycle, resolving startup database initialization errors, and fixing data seeding reliability issues reported during InterSystems Open Exchange verification.

Key Changes & Enhancements:
Build-Time FHIR Provisioning: Moved FHIR server installation (/fhir/r4) and recursive class compilation directly into the docker build phase (iris.script). This ensures all schema tables, vector indices, and custom classes are baked directly into the container image, resolving 404 endpoint errors and startup race conditions.

REST Output Stream Protection: Implemented a quiet mode (pQuiet parameter) in the database Setup() method. This suppresses terminal writes when database initialization is triggered from REST controllers, preventing HTTP response pollution and resolving client-side JSON parsing errors (Expecting value: line 1 column 1) during data seeding.

Programmatic Namespace Safety: Wrapped the sample data loader (LoadSampleData()) in a scoped in-memory namespace switch to INTEROP. This resolves the ServiceIdIdx registry lookup failure if the CLI seeding command is executed from a default namespace (e.g. USER).

Robust Container Start Fallbacks: Updated the runtime initialization script (init_iris.sh) to recursively compile all custom classes (including ClaimAudit.Data and ClaimAudit.FHIR) using LoadDir, ensuring correct compilation ordering when deploying with empty persistent volumes.

ZPM Package Alignment: Correctly registered ClaimAudit.Data.PKG in module.xml to track and compile the GraphStore class within ZPM.

Expanded Diagnostic Wiki: Added new troubleshooting guides to the project wiki detailing resolution steps for ServiceIdIdx failures, missing database tables, and REST output pollution.

What's Changed

• fix: resolve Gradio UI overlapping and uneven layout issues by @jrpereirajr in https://github.com/musketeers-br/TriageAide/pull/17

Full Changelog: https://github.com/musketeers-br/TriageAide/compare/1.0.1...1.0.2

Main changes:

• improvements on instructions presented in README.md
• fix agent memory which was leading to repetitions

What's Changed

• Fix/gradio error handling by @jrpereirajr in https://github.com/musketeers-br/TriageAide/pull/1
• feat: add LangSmith observability for agent tracing by @jrpereirajr in https://github.com/musketeers-br/TriageAide/pull/2
• feat: separate IRIS FHIR server and triage app into independent Docke… by @jrpereirajr in https://github.com/musketeers-br/TriageAide/pull/3
• feat: add agent observability to Gradio UI with trace panel by @jrpereirajr in https://github.com/musketeers-br/TriageAide/pull/4
• docs: sync all documentation with codebase after Docker separation, L… by @jrpereirajr in https://github.com/musketeers-br/TriageAide/pull/5
• feat: improvements on conversation examples by @jrpereirajr in https://github.com/musketeers-br/TriageAide/pull/7
• Feat/voice bridge fix and docs by @jrpereirajr in https://github.com/musketeers-br/TriageAide/pull/8
• feat: integrate ElevenLabs voice I/O with bilingual support (pt-BR + en) by @diashenrique in https://github.com/musketeers-br/TriageAide/pull/6
• fix: auto language detection by @jrpereirajr in https://github.com/musketeers-br/TriageAide/pull/9
• feat: add ENABLE_VOICE_UI flag to hide ElevenLabs UI for MVP, mark vo… by @jrpereirajr in https://github.com/musketeers-br/TriageAide/pull/10
• Conversation fixes by @jrpereirajr in https://github.com/musketeers-br/TriageAide/pull/11
• Conversation fixes by @jrpereirajr in https://github.com/musketeers-br/TriageAide/pull/12
• Conversation fixes by @jrpereirajr in https://github.com/musketeers-br/TriageAide/pull/13
• fix: images for agent diagrams by @jrpereirajr in https://github.com/musketeers-br/TriageAide/pull/14
• Conversation fixes by @jrpereirajr in https://github.com/musketeers-br/TriageAide/pull/15
• Conversation fixes by @jrpereirajr in https://github.com/musketeers-br/TriageAide/pull/16

New Contributors

• @jrpereirajr made their first contribution in https://github.com/musketeers-br/TriageAide/pull/1
• @diashenrique made their first contribution in https://github.com/musketeers-br/TriageAide/pull/6

Full Changelog: https://github.com/musketeers-br/TriageAide/commits/1.0.1

Updated README with links to articles and videos

• Updated all databases to latest versions: IRIS 2026.1, PostgreSQL 18, MySQL 9.7.0, SQL Server 2025
• Refreshed benchmark results with new performance comparisons
• Fixed docker-compose worker auto-registration and startup issues

Link to YouTube demo included in description

Added public access urls for demo and wiki.

Updated README

v1.0.0 — Initial release

A clinician dashboard with AI triage questions that automate pre-checkup intake, built on InterSystems IRIS for Health.

• Patient intake interview saved as a FHIR QuestionnaireResponse
• AI agent reads the patient's FHIR record and retrieves matching clinical guidelines via IRIS Vector Search (%Embedding.OpenAI + VECTOR_COSINE)
• Produces a cited triage decision (self-care / see-GP / urgent-care / ED) with a deterministic red-flag safety gate
• Writes the full cascade back to FHIR: Encounter, SNOMED-coded ServiceRequest, LOINC/SNOMED Observations, and a Communication alert on escalation
• Clinician worklist to review cases, with FHIR write-back on acknowledgement
• Runs through an IRIS Interoperability production (every triage visible in Visual Trace)
• One-command docker compose up — boots IRIS, the agent, and the UI, and self-seeds demo data

Initial Release

Initial Release

Initial Release

Initial Release

Initial Release

Initial Release

Initial Release