Final-year Computer Science student at King's College London (Predicted First Class). I work close to the metal - systems programming, low-latency infrastructure, and understanding exactly why things break.
My interests sit at the intersection of systems and security: compilers, memory safety, and how low-level vulnerabilities actually manifest. I built a coverage-guided fuzzing framework targeting Poly/ML (the runtime at the core of Isabelle/HOL's trusted computing base), finding UBSan and memory safety bugs in ARM64-specific compiler code.
Trust Me, I am a Verifier! (Or should you?) - Fuzzing the Poly/ML Compiler · C, Standard ML, AFL++, LLVM, ASan/UBSan
- First systematic coverage-guided fuzzing framework for Poly/ML (Isabelle/HOL's trusted computing base) on ARM64
- Direct AFL++ binary fuzzing with LLVM LTO PCGUARD instrumentation; ASan/UBSan layered at runtime to avoid bootstrap failures; CMPLOG and rare power schedule added for Phase 2
- 72 curated SML seeds + Isabelle/HOL corpus; two-phase lexer/parser strategy with afl-cmin minimisation and evolved corpus handoff between phases
- Found 4 reliability findings in upstream Poly/ML: ARM64-specific UBSan overflow in arm64.cpp:246, lexer OOM on pathological float literals, and 3 SIGSEGV crashes from a type safety defect in overloading resolution
- Overloading defect confirmed and fixed upstream in cf7b84a following this bug report; two fixes proposed and validated locally
Crux - Optimal Rubik's Cube Solver · C++17, CMake
- Guarantees minimum-move solutions (≤20 moves, God's Number) via IDA* with pattern databases
- 88M-state corner pattern database stored as 4-bit nibbles (~42 MB); three 6-edge partial DBs built via BFS
- O(1) heuristic evaluation via coordinate move tables - no per-node arithmetic on the search hot path
- Parallel search across 18 root moves with atomic abort; 12-move scrambles under 30ms, 32-test suite validates optimality
DCache - High-performance distributed in-memory cache · Go, Docker, Prometheus
- 256-shard concurrent map achieving 50M+ ops/sec with sub-25ns GET latency; 80+ Redis-compatible commands
- AOF persistence (buffered-channel writer, configurable fsync) + CRC-32C binary snapshots for crash recovery
- Async master-slave replication via PSYNC with bounded ring-buffer backlog and TCP connection hijacking
- Prometheus observability with per-command latency histograms; Docker Compose stack with Grafana
Real-Time Market Data Simulator - Low-latency market data streaming engine · Python, asyncio, NumPy
- 1.4M+ ticks/s raw generation; 127k+ msg/s end-to-end across 10 subscribers with sub-200µs p99 latency and zero loss
- GBM price dynamics; asyncio fan-out with per-subscriber queue isolation and backpressure handling
- p50/p95/p99/p99.9 percentile tracking via NumPy vectorised operations, decoupled from the generation hot path
- 53-test suite covering GBM statistical properties, backpressure behaviour, and high-load end-to-end scenarios
Auteur - Personal cinema intelligence engine · Python, FastAPI, React, ChromaDB, SQLite
- Local RAG pipeline: sentence-transformers embeddings in ChromaDB; semantic retrieval drives a streaming conversational AI (Ollama / Anthropic)
- Progressive taste profiling: fast DB-only stats endpoint + async LLM endpoint for cinematic identity and hidden pattern detection
- Full-stack: FastAPI + SQLite backend, React + Tailwind frontend with real-time streaming chat and filter/sort history browser
Software Engineer Intern - The Kusp Hub (June - September 2025)
- Architected AI-powered career discovery platform with 97%+ CV skill extraction accuracy using LLM structured outputs (Groq API) and coordinate-based multi-column PDF reconstruction with PyMuPDF
- Engineered two-stage semantic job matcher combining sentence-transformer bi-encoder retrieval with category-weighted reranking; reduced pipeline latency by 95% (75s → 2s) through offline vector pre-computation
- Built full-stack Flask application with async processing, confidence scoring and semantic match explanations; deployed to production on Render with Gunicorn
- Developed work credits generator using custom LLM prompts to categorise and surface candidates' creative industry experience as a portfolio PDF
Languages: Python, C/C++, Java, Go, Scala, JavaScript, SQL, Standard ML
Systems & Tools: Docker, Bash, AFL++, LLVM, ASan/UBSan, asyncio, NumPy, Redis, Prometheus, CMake
Frameworks: Flask, Django, React, FastAPI