Every project below has something running behind the link. The screenshots are captures of the live pages, not mockups.
A calm optical puzzle game for iPhone. Guide beams through mirrors, prisms, splitters and portals to ignite every receiver: 40+ handcrafted levels, star-based progression, no timers. Free to play, with landing, privacy, and support pages on this domain.
A puzzle is only calm when the optics, the levels, and the store around them all agree.
Separates sustainable dryland greening from vegetation quietly draining a finite water store, across five regions in China, Saudi Arabia, and Türkiye. Every modelling task is scored against its own dummy baseline.
Every result is scored against its own dummy baseline, so the answer survives review.
An agentic compliance operating system that turns regulation into auditable workflows. Rules-first decision engine, multi-tenant Postgres, and LLMs used only where the text is genuinely ambiguous. First plugin covers the EU AI Act.
Infrastructure feeds the models, and the models feed a product that has to answer to regulators. Each layer sets constraints on the one after it.
A declarative, database-driven multi-agent runtime. Multi-provider LLM routing with failover, two-tier semantic caching, and cost, latency, and quality observability, shipped as one static binary next to Postgres.
Multi-agent work collapses into one static binary, with Postgres keeping the books.
Maps each checklist question to the clauses that actually govern it and returns evidence-backed answers in minutes. Runs in the cloud or entirely on a laptop.
A checklist question goes in; a clause-backed answer comes out, on the cloud or a laptop.
A cross-platform product comparison app that searches several marketplaces at once and ranks results with AI-assisted recommendations. An iOS release is planned.
Several marketplaces, one query, recommendations ranked by a model, not by ad spend.
Internal monitoring for AWS Step Functions and Batch workloads, giving engineers one place to see what is queued, running, and failing.
One pane of glass for what is queued, running, and failing across the pipeline.
A Go tournament service with a Redis-backed live leaderboard, built during the BTS internship and deployed to AWS and Minikube. The project page runs the scoring rules in the browser.
Scoring rules run live in the browser while Redis keeps the leaderboard honest.
Deployed an open-source enterprise generative AI platform to AWS with Terraform. Evaluated as a lecture assistant at Lund University, where the running cost ruled it out.
Deployed, evaluated in real lectures, and shut down on the numbers; the cost made the call.
New projects, research posts, and short notes on what shipped. Roughly monthly, never more.
Double opt-in: your address is only stored after you confirm the link I email you. No spam, no sharing, one click to leave; every email carries an unsubscribe link.
