I build AI systems that know what they don’t know.

Open-source FHIR R4 and compliance SDK for India's ABDM health network. Published to PyPI, Apache 2.0 licensed.

The problem. A clinic in rural Andhra Pradesh wants to send a discharge summary to a hospital in the next district. Between them sits ABDM: ABHA identity, consent artefacts, encrypted exchange, and a gateway with strict async callback contracts. A two-person team can't build all of that before they build the thing they actually set out to build.

What I built. A Python SDK covering the whole path: ABHA identity, HIP/HIU consent and data-exchange flows, FHIR R4 bundle builders, ECDH X25519 key exchange with AES-GCM encryption, gateway resilience through retries and circuit breakers, clinical templates across 12 medical domains, WhatsApp and AI-assisted clinical workflow helpers, and a compliance engine with rule packs for India, Australia, the US and the UK. India's ABDM adapter is the one that talks to a live network; the Australia, UK and US adapters expose identifier and formulary metadata and raise NotImplementedError on network calls. v1.0.0-alpha.3 on PyPI, Apache 2.0, 128 tests, CI on Python 3.10–3.12 with pytest, ruff, bandit and pip-audit.

One decision. Compliance returns blockers, not warnings. The compliance engine is a pre-flight check: purpose, lawful basis, encryption, residency, minimum-necessary sharing. Each failed check comes back as a blocker with a source URL, and the result only passes when the blocker list is empty. It would have been friendlier to emit warnings and let the caller proceed, but the failure mode you're designing against is an engineer shipping on a deadline, and a warning is something you ship past. Country rules live in swappable policy packs, so the same call site is legal in India and refuses in Australia when it should be.

Cross-game LLM deception benchmark: five social deduction games behind one engine interface, with structured decision traces.

The problem. If a model is persuasive under incomplete information, you want to know that before you put it somewhere consequential. Most evaluations ask a model to answer honestly. Almost none put it in a game where deceiving another player is the winning move and then measure what it does.

What I built. Skull, Coup, The Resistance, One Night Ultimate Werewolf and Secret Hitler implemented as pure state machines behind a single interface; 1000/1000 randomised stress tests, 200 seeded games per engine. On top of that: Deception ELO ratings, a 10-axis capability taxonomy, an aiohttp WebSocket server, SQLite run persistence, and a four-screen React dashboard for run setup, live monitoring, analytics, and replay. 15 models across 5 providers are wired in.

One decision. Score the public rationale, not the chain-of-thought. Every model turn writes a structured decision trace: stated rationale, confidence, suspected deceivers, tactic, risk assessment, latency, provider. The dashboard deliberately surfaces the public rationale and not hidden reasoning. A deception benchmark that scores a model on its private thoughts is measuring the wrong object: deception is made of what the other players can observe. Each engine also carries a simulation-fidelity label, so a number never implies more game than the engine actually implements.

Scope, honestly. Real-API runs so far are small: one game per level, two models. The larger multi-model run was an offline deterministic stand-in, and the exports say so. What exists is benchmark infrastructure, not findings.

AI triage and referral platform for the Lake Macquarie & Newcastle Suicide Prevention Network. Built with a teammate at NGM Group's hackathon, May 2026.

The problem. Someone reaches out to a small suicide-prevention network at 11pm. The staff who could help are asleep, the intake form is a PDF, and whoever picks it up tomorrow has no way to tell whether this person needs a callback next week or a phone call right now.

What I built. A 24/7 intake chatbot, a QR-based referral portal for frontline staff, and a dashboard that gave the organisation its first analytics. Rule-based risk scoring routes each referral to staff by the training level it requires, across 14 programs, backed by a five-layer escalation protocol and 35+ documented edge cases. FastAPI, Azure AI Language, React. LMNSPN's CEO has since met with us about taking it into production.

One decision. Crisis numbers survive the backend. Lifeline's number is hardcoded in static HTML. If the API is down, the model is rate-limited, or the deploy is broken, the page still shows a human being to call. Everything else in the system is best-effort; that one element is not permitted to depend on anything. The same instinct drove the risk engine: it's rules, not a model, because when it's wrong we need to be able to read why, and borderline cases escalate rather than resolve.

Cultural research workspace that labels every AI claim by how far you should trust it. Built solo in about five hours at the OpenAI Codex Hackathon, Sydney.

The problem. A researcher asking about a 14th-century trade route gets back a fluent paragraph with no way to tell which sentence came out of an archive and which one the model inferred. The fluency is the problem: it makes an inference and a citation look identical.

What I built. One query fans out to ten analysis surfaces and six live archive APIs, producing a knowledge graph, geographic transmission flows, image analysis with bounding boxes, and PDF/PowerPoint export. Next.js 14, TypeScript, GPT-4.1.

One decision. Every claim carries its epistemic status. Output is labelled fact, context, hypothesis, or needs-expert-review, and the label travels with the claim all the way into the exported deck. It makes the system look less confident than a plain summary would, which is the point: a researcher can act on a fact and go verify a hypothesis, but a paragraph that blends the two can't be checked at all.