How it works
How these projects work
You don't need to understand AI to understand what these applications do. For each one: what goes in, what happens, and what comes out.
Healthcare interoperability + AI workflow automation
Healthcare Intelligence Platform
A platform that receives healthcare data, checks it, cleans it into one consistent format, and uses an AI agent to find and explain relevant information — always pointing to the evidence it used.
What goes in
- Healthcare messages from other systems (HL7 v2 admissions, discharges, lab results)
- Clinical documents (C-CDA visit summaries) and FHIR records
- A question in plain English, such as “Does this patient need a follow-up?”
What happens
- Each message is checked for missing, invalid, or duplicate information
- Everything is translated into one consistent format (FHIR) and given a quality score
- An AI agent picks the right lookups, searches notes, and gathers evidence
- The answer is checked against the evidence before it is shown
What comes out
- Clean, consistent patient records with a data-quality report
- A short answer that cites the exact records and notes it relied on
- A trace showing every step the agent took
See every stepThe full flow, one or two sentences per step — with optional technical notes.
- 01
Healthcare data arrives
Messages and documents come in from other systems — admissions, discharges, lab results, and visit summaries.
Technical implementation: HL7 v2.5.1 (ADT^A01/A03/A04/A08, ORU^R01), C-CDA R2.1, FHIR R4 bundles via REST ingestion endpoints.
- 02
Validation
Each message is checked for problems: missing fields, invalid codes, broken references, or duplicates. Bad records are set aside instead of silently stored.
Technical implementation: Staged pipeline: parse → syntax → schema → quality. LOINC mod-10 and SNOMED Verhoeff checks, MSH-10 duplicate detection, quarantine on error, HL7 ACK (AA/AE/AR).
- 03
Normalization
Different systems describe the same thing differently. The platform lines up codes, dates, and patient identities so everything matches.
Technical implementation: Code-system mapping, timestamp normalization, lightweight patient matching (MPI-lite), ADT merge handling.
- 04
FHIR
Everything is stored in one modern, standard format, and each record gets a 0–100 data-quality score.
Technical implementation: Canonical FHIR R4 resources in PostgreSQL; custom declarative validator for 10 resource types.
- 05
AI Agent
When you ask a question, the agent first screens it for safety, works out which patient and what kind of question it is, then plans which tools to use.
Technical implementation: LangGraph graph: safety_screen → classify_intent → identify_patient → plan_tools → execute_tools → … → finalize.
- 06
Tools + RAG
The agent looks up encounters, medications, labs, and appointments, and searches clinical notes for relevant passages.
Technical implementation: 9 typed clinical tools + hybrid retrieval (vector + keyword + section prior) over pgvector, with a relevance floor.
- 07
Evidence
Everything the agent found is collected as evidence, and the answer is checked against it. Unsupported sentences are rejected.
Technical implementation: Grounding check: each claim must cite evidence, share ≥50% of terms, and match every number and date. LLM answers under 0.7 fall back to an extractive answer.
- 08
Answer
You get a short, plain answer with citations — and a trace you can open to see exactly how it was produced.
Technical implementation: Server-sent events streaming, per-run traces with tool calls, latency and token usage.
AI-powered personal information and productivity assistant
PersonalOps AI
A personal AI assistant that reads your documents, pulls out the useful details, answers questions with evidence, and helps keep track of bills, subscriptions, warranties, and deadlines — but never changes anything without your approval.
What goes in
- Personal documents: receipts, bills, warranties, subscription confirmations, insurance and travel documents (PDF, Word, text, CSV, JSON)
- A question in plain English, such as “When does my laptop warranty end?”
What happens
- Each document is read, sorted into a category, and key details are pulled out
- Those details become organised records — purchases, bills, subscriptions, warranties, deadlines
- The assistant searches the records and documents and collects evidence
- If action is needed, it drafts a task or reminder and waits for your approval
What comes out
- An answer with the source documents and a confidence level
- Upcoming bills, renewals, and expiring warranties in one place
- Tasks and reminders — only the ones you approved
See every stepThe full flow, one or two sentences per step — with optional technical notes.
- 01
Upload document
You add a receipt, bill, warranty, or similar file. The file is checked for type and size before it's accepted.
Technical implementation: Upload validation (type, 15 MB limit), local or S3 storage.
- 02
Read document
The system reads the text out of the file.
Technical implementation: pypdf for PDF, python-docx for Word; CSV/JSON/TXT natively. Scanned images are flagged — OCR is not enabled.
- 03
Extract information
It works out what kind of document it is and pulls out details like merchant, amount, due date, or warranty end date.
Technical implementation: Deterministic weighted-keyword classification (user can override) + label/regex field extraction.
- 04
Store information
Those details are saved as organised records, each linked back to the original document.
Technical implementation: PostgreSQL tables for Purchase, Bill, Subscription, Warranty, Deadline — linked to their source Document.
- 05
RAG search
The document text is split into small passages and indexed so the assistant can find the right passage later.
Technical implementation: Line-aware chunking (800 chars / 120 overlap), embeddings in pgvector, hybrid lexical + vector scoring.
- 06
AI agent
When you ask a question, the agent works out what you're asking about and which lookups it needs, then runs them.
Technical implementation: LangGraph: intent detection → context detection → tool selection → run tools → RAG retrieval → evidence validation → answer → action detection.
- 07
Evidence
Every answer comes with the documents and records it relied on, plus a confidence level. If there's no evidence, the assistant says so instead of guessing.
Technical implementation: Evidence items typed as extracted / database / user-provided / inference, scored high-medium-low; refusal when no evidence.
- 08
Suggested action
If it notices something important — say, a subscription renewing next week — it drafts a task or reminder.
Technical implementation: Mutating tools (create_task, update_task, create_reminder) only create pending ActionProposal rows.
- 09
Your approval
You review the suggestion and approve or reject it. Nothing changes until you click approve.
Technical implementation: Run ends in awaiting_approval; the action executes only through the explicit approve endpoint.
- 10
Task / reminder
Approved suggestions become tasks or reminders, and upcoming ones appear on your dashboard.
Technical implementation: Background worker marks reminders as due on a polling interval.
Glossary
What the words mean
The technical terms used across this site, in ordinary language.
- AI agent
- An AI that works through a set of steps and uses tools — like a search or a database lookup — instead of answering from memory.
- RAG (retrieval-augmented generation)
- “Look it up, then answer.” The system finds relevant passages first, and the answer is built from them.
- Evidence / citations
- The specific records or passages an answer is based on, shown so you can check them.
- Tool
- A small, specific function the agent can call — for example “find this patient's appointments.”
- Human-in-the-loop
- A person approves important actions before they happen.
- Validation
- Checking data for missing, invalid or inconsistent information before it's used.
- LLM (large language model)
- The kind of AI that writes text. In both projects it's optional — they also work without one.
- LangGraph
- The software library used to define the agent's step-by-step workflow.