I like taking things apart until I understand exactly how they work - these days that's mostly LLMs and AI systems, with some compilers and distributed systems in the mix.
Fuzzing the Poly/ML Compiler: C · LLVM · AFL++
First systematic coverage-guided fuzzer for Poly/ML (Isabelle/HOL's trusted computing base) on ARM64. Dissertation scored 84%. Found 3 defects, including one confirmed and fixed upstream, and an ARM64 overflow I diagnosed with a proposed fix.
DCache: Go · Redis protocol
Redis-compatible distributed cache. 50M+ ops/sec, sub-25ns GET latency across a 256-shard concurrent map.
Crux: C++ · CMake
Optimal Rubik's Cube solver. Guarantees ≤20-move solutions via IDA* with pattern databases; solves 12-move scrambles in under 30ms.
Real-Time Market Data Simulator: Python · asyncio
Streaming market data engine. 1.4M+ ticks/sec, sub-200μs p99 latency, zero message loss.
Auteur: Python · FastAPI · React
Local RAG app that builds a conversational "cinematic identity" profile from personal watch history.
Software Engineer Intern, The Kusp Hub - Jun-Sep 2025 Built an AI-powered career discovery platform: LLM-based CV extraction and a semantic job-matching pipeline, cutting latency from 75s to under 2s.