Sole engineer on most projects, owning each from PRD and prototyping through infrastructure, deployment, and client demos.

A platform that turns a plain-English prompt ("a diamond sword that shoots lightning") into a compiled, installable Minecraft mod through an agentic loop with no human in it: an agent writes Fabric/Java against retrieved API context, a sandboxed Docker + Gradle container compiles it, and build errors feed back until it passes. I built the full product around it, and reverse-verified Minecraft 26.1's newly unobfuscated API from the compiled JARs after official mappings were discontinued.

A coverage-guided fuzzing agent that feeds a function's uncovered branches back to an LLM each iteration to target unhit paths, instead of mutating blindly like AFL or LibFuzzer. On a benchmark tokenizer it hit 98.3% branch coverage vs 90.7% for random fuzzing and found 4 crashes the baseline missed, using under half the inputs.

A Java web search engine running on distributed infrastructure I built from scratch: a sharded, consistent-hashing key-value store and a Spark-style compute engine with an RDD interface. On top of it, a crawler over 300K+ pages, a distributed inverted index, iterative PageRank, and multi-signal ranking that serves sub-50ms queries.

A from-scratch voxel engine in C++ and OpenGL (3-person team) where I owned rendering and generation: chunked terrain that streams in and out around the player in real time, a full GLSL shader stack with a procedural analytic day/night sky, and rivers and asset scatter from a stochastic L-system.

A full-stack system that rethinks the post-stay hotel review: instead of a generic form, it mines a property's existing reviews for what is missing, stale, or contradicted, then asks one or two targeted follow-ups. A deterministic gap-scoring engine does the analysis offline and an LLM only phrases the questions, at roughly a tenth of a cent per insight. Placed 3rd of 65 teams at the 2026 Wharton × Expedia Hack-AI-thon.

My research targets the 'vigilance gap': a model holds a safety-critical fact a user shared earlier (say, a latex allergy) but fails to act on it once the question stops pointing back to it ("banana or avocado smoothie?").

I approach it as a memory problem. Most memory systems are built to fetch what a user is asking about, while vigilance needs the opposite: surfacing a fact they are not asking about but urgently need.

A memory system I designed for safety rather than general recall raised success on a safety benchmark from 48% to 87%, while cutting false alarms from 11% to 8%, below the base model. I also built the parallelized, cost-tracked pipeline that runs these evaluations at scale.