Projects

Everything I've built that's worth showing, newest first within each group. Anything with a page attached has screenshots, numbers and a short note on what I'd do differently.

Companies

Transmute2026 —A compiler that turns a production AI agent into software you actually own. Parity with GPT-5.2 on 180 sealed tasks, higher repeatability, and about 18 times lower cost.Insora2025–26A secure AI workspace for European SMEs. I co-founded it, led engineering for a year, had 50+ discovery conversations across about 25 SMEs, and delivered a client pilot in six days.

Open source & tools

A personal AI computer2026 —A personal assistant that runs on my Mac, talks through Telegram, and can change its own tools and code. Live tests covered self-change, restart, rollback, approvals and replay.Parity2026 —A Stripe Billing simulator that checks itself against the real sandbox. An LLM agent's outcomes agreed on 79/80 billing tasks; API response fidelity is measured separately.Slim2026A CLI that deletes npm dependencies. It figures out how you use a package, writes a replacement, fuzzes it against the original, and opens the PR with the evidence attached.CAMEL-AI contribution ↗2025Merged PR #3941 into the 17k-star multi-agent framework: Azure OpenAI Responses API support, sync and async execution, structured outputs, and tests.

Research

Agent harnesses for cloud infrastructure2026 —Better tools and feedback for small models writing Terraform. Qwen 27B's policy pass rate went from 18% to 31% on 50 held-out tasks, reaching the earlier Max baseline.Verified self-play for code optimization2026 —Models search for LLVM optimizations, checked with alive2 and real hardware timing. Two fragments from Google's re2 ran 14–16% faster than -O3. With Professor Gabriel Poesia.Self-play for PyTorch optimization2026 —Extending the verified code-optimization work to PyTorch with Professor Gabriel Poesia. The next question is whether models can learn from proposing and solving optimization tasks.Project Vera2026 —World-model emulators of any cloud, used as training environments for infrastructure agents. I'm building the harness for cloud-agent RL. With Professor Ang Chen.AER — durable recovery for agents2026Lets crashed agents resume without repeating side effects. 48% less work after recovery than checkpoint-only, and it recovered in every trial where the baseline failed.RL-learned KV retention2025A DPO-trained controller that decides what to evict from the KV cache. Up to 1.37 perplexity better than sliding-window attention on PG-19, at under 0.1% latency overhead.Efficient, faithful RAG2025REFRAG-inspired adaptive context selection on FEVER. 88.2% accuracy with 1.5 sentences on average, compared with 85.6% for full context at 11.1 sentences.Editing constraint models with LLMs2025My first-author paper at CP 2025 in Glasgow. I defined the task, built the benchmark, and reached 95.8% compile accuracy with GPT-4o.Mamba vs. Transformers for satellite super-resolution2025My undergraduate thesis at St Andrews. SwinIR and MambaIR built from scratch in PyTorch and Triton on a cross-sensor dataset I put together; 20.5 dB PSNR.

Shipped at work

Eval and optimization framework at MAMA AI2026An evaluation and prompt/program optimization harness for their production LLM pipelines. Multimodal extraction accuracy went from 77% to 87%, and I packaged it as a config-driven tool other teams can reuse.

AskYourData at Blindspot AI ↗2023Enterprise RAG for internal documents. I improved retrieval accuracy by over 70% through chunking, embedding, index and prompt changes, and built real-time web ingestion in Scala and Azure Functions.

Weekends & hackathons

PreferenceGraph ↗2026A writing-style preference learner built in 24 hours at JacHacks: Elo ranking, active-learning comparison selection (BALD), and loss-aversion weighting.Knowl2026A Cursor-style workspace that logs every prompt and draft so evaluators can grade the process rather than the output. Built in 24 hours at the Michigan AI Hackathon, and it's live.RadioGuessr2024Guess the country from a live radio stream on a 3D globe. Built in 24 hours with a team of four that I led, and we came second.

Earlier

Cognate2024A Notion-style workspace with AI-generated study notes structured around Bloom's taxonomy. Custom block editor, databases, Convex.

StudyBuddy ↗2024A peer study-partner platform. I validated it with a no-code prototype first, then rebuilt it as a real web app, and it won funding from FastStart Digital Launchpad.

Accessible Wikidata stadium visualizer2024A team project at St Andrews where I led the front end and visualization, designed screen-reader first.

STACS website2024The site and sponsor CRM for the St Andrews Computing Society, where I led the developers subcommittee and ran workshops on things like vision transformers and post-quantum cryptography.

Sumitify2023A landing page, pricing and scorecard quiz for a productized AI services studio idea.

CoPilot3602023A product design exploration of AI co-pilots for product teams. The design itself was the artifact.

Robot maze game for my CS club2021A C++ game where students program a robot through procedurally generated mazes. I built it for the CS club I started in high school, which grew to 35+ students across several schools.

New projects get added as rows here. Only the ones worth a full write-up get their own page.