Systems for complex domains.

Overview

A connected, machine-readable map of gerontology and healthy-aging research. It exposes queryable REST and MCP APIs with semantic, natural-language search, and keeps every claim traceable to its evidence.

Why I built it

Research on aging is vast and fragmented. I wanted a knowledge base that both people and AI agents could query and extend — without losing the link back to the underlying citations.

What it does

• Serves a structured knowledge graph over REST + MCP, with pgvector semantic search.
• Runs Claude curator agents that expand the graph with real citations (PubMed / DOI).
• Uses a reviewer agent to verify citations and promote verified claims to “curated”.
• Exposes a token-gated write path with an admin curation UI.

Architecture & approach

An ontology-driven knowledge model defines domains, node types, relationships, and an evidence model. A Postgres schema with migrations is paired with a JSON seed graph as the single source of truth; a validator enforces schema compliance, and CLI tools handle seeding and embedding generation. It runs against ephemeral Postgres for offline testing or a persistent Neon backend, deployed on Cloud Run with agents on Cloud Scheduler.

Key technical challenges

• Keeping AI-generated additions trustworthy through citation verification and reviewer promotion.
• Maintaining end-to-end evidence traceability as the graph grows.
• Tuning semantic retrieval quality with pgvector embeddings.
• Aligning with health standards (FHIR, SNOMED CT, open vocabularies).

Overview

Both a software platform and a piece of research infrastructure: it continuously observes aging-policy systems and measures outcomes against consistent metrics, tracking policy lifecycles, budgets, target populations, and indicators over time.

Why I built it

Policy across jurisdictions is difficult to compare consistently. I wanted an independent, reproducible instrument with documented methodology, rather than a one-off report.

What it does — five modules

• Policy Library — jurisdiction-aware catalog with full text and AI summaries.
• Data Hub — versioned, lineage-tracked open-data ingestion.
• Indicators / HAPI — a six-domain composite index across health, independence, social participation, financial security, care access, and digital inclusion.
• Policy Analytics — association and causal-inference analysis with explicit rigor.
• AI Research Assistant — a retrieval-augmented agent for evidence packs and literature reviews.

Architecture & approach

A monorepo separates concerns cleanly: a Next.js web app for dashboards, a Python pipeline for ingestion, ETL, and statistics, a Postgres database with schema and migrations, a shared contracts package for enums, and a documented design whitepaper. The architecture is multi-tenant-ready.

Key technical challenges

• Data lineage and versioning so results stay reproducible.
• Methodological rigor in a theory-anchored composite index.
• Being explicit about association versus causal inference.
• Modeling multiple jurisdictions under one consistent schema.

Overview

A structured knowledge base and web application that presents traditional acupoint knowledge alongside modern anatomy, mechanism notes, and clinical evidence grading — 161 acupoints across 14 meridians, in four languages.

Why I built it

Traditional knowledge is often presented without provenance or modern context. I wanted a structured, multilingual reference where each recommendation carries an explicit evidence grade and red-flag reminders.

What it does

• Interactive 2D/3D body viewer with per-point markers.
• “Symptom → points” reverse lookup with full-text search.
• A–D evidence grading per symptom, with mechanism and safety notes.
• Massage logging with personal “what worked for me” aggregation.

Architecture & approach

A data layer holds acupoints, meridians, symptom mappings, evidence grades, SVG diagrams, and a 3D model; a build pipeline generates JSON, populates SQLite, and handles internationalization; a FastAPI layer serves reference data, search, user records, feedback, and auth; and a single-page front end ties it together with the body viewer and encyclopedia.

Key technical challenges

• Modeling structured knowledge with per-symptom evidence grading.
• Fast full-text search via SQLite FTS5.
• Interactive 3D anatomy with Three.js, kept lightweight and PWA-ready.
• Four-language internationalization and anti-abuse on anonymous feedback.

Overview

Models how a research programme for digital interventions in healthy aging could be architected to warrant trust — through transparency, participant agency, and a rigorous separation of research concerns. It runs on synthetic data only and is explicitly non-empirical.

Why I'm building it

I wanted to see whether a trust-first model for sensitive health research holds up under engineering — whether an invariant survives a case designed to break it, and whether the governance stays coherent in code rather than only on paper.

What it does

• Participant-controlled consent and permissions, granular per person.
• Participant-authored life stories, with supporter contributions marked distinctly.
• Research-chain separation: dataset → locked version → analysis → output → interpretation.
• Audit logging of every permission decision.

Architecture & approach

A TypeScript / NestJS modular monolith across eighteen domain modules with statically enforced boundaries, split into API, worker, and scheduler processes. Durable messaging uses a transactional-outbox / idempotent-inbox pattern over pg-boss, and modules talk only through their contracts. It integrates the Aging Knowledge Graph over MCP.

Key technical challenges

• An Effective Permission engine — a pure function over role, relationship, consent, purpose, context, and resource state.
• A “protected existence” principle: refusals never reveal that a record exists.
• Seventeen AI actions refused by name; automation cannot confirm safety signals.
• Keeping eighteen modules independent under enforced dependency boundaries.