Engineer working in AI systems, backend infrastructure, and applied research
trust-gate TypeScript
Trajectory-aware regression gate for autonomous coding agents. Open-source and self-hostable: Fastify API, Next.js dashboard, MCP server via xmcp, BullMQ/Redis workers, and Postgres with Drizzle. Ships a reusable GitHub Actions check (trust-gate-demo) to gate agent-generated PRs end-to-end.
agent-arena Python
Benchmarking platform for autonomous coding agents. Define benchmarks and tasks, run agents in isolated git workspaces, and stream every event live over WebSocket. Trace viewer replays full executions — tool calls, file diffs, stdout/stderr, and token usage with cost breakdown. Claude Code adapter first, extensible to any agent.
chainforge Go
Provider-agnostic AI agent framework with zero external dependencies. Sequential, parallel, and router-based orchestration. MCP (HTTP + stdio), vector memory (Qdrant, PostgreSQL, Redis), tool memoization, Prometheus metrics, and OpenTelemetry tracing. Swap Anthropic, OpenAI, Ollama, or Gemini with one line.
dense-retrieval-ft Python
Fine-tuned BGE-M3 with LoRA for domain-specific dense retrieval on a single consumer GPU. Synthetic query generation and hard-negative mining pipeline, contrastive training, and retrieval evaluation against the base model.
etip Python · TypeScript
Enterprise Talent Intelligence Platform. Ingests GitHub and Jira activity, infers skills against the ESCO taxonomy, and ranks candidates via vector similarity + cross-encoder reranking + LLM explanations. Multi-tenant with PostgreSQL Row-Level Security, pgvector, Celery workers, and Qdrant.
websearch-ng Python · TypeScript
Self-hosted AI search engine. Parallel web search, Jina extraction, semantic reranking, and cited answers streamed token-by-token over SSE. Two modes: fast search (3 queries) and deep research (7 queries + gap analysis + structured report). Multi-provider via LiteLLM.
gpt-sota-opt Python
Production GPT implementation with GQA (25–30% faster inference, 66% smaller KV cache), FlashAttention-2, RMSNorm, SwiGLU, and RoPE. Scales from 10M to 500M parameters. 6,000–7,000 tokens/sec on RTX 5070 with BF16 and gradient checkpointing.
turboquant-implementation Python
TurboQuant KV cache compression — 4× size reduction with near-zero quality loss. RHT + Lloyd-Max quantization encoded at 3–4 bits. Custom Triton GPU kernels: ~9× faster quantization, ~5× faster dequantization. Validated on Llama-3.2-3B: +2.2% perplexity, 100% needle accuracy.
rl-framework-qwen3.5-2B Python
Full SFT + GRPO reinforcement learning pipeline for reasoning on Qwen3.5-2B. QLoRA 4-bit quantization fits on 12GB VRAM. Self-improvement loop: generates reasoning traces, filters top performers, augments the dataset, and repeats. Evaluated on math and code benchmarks with structured <think> chain-of-thought.
autonomous-sae-researcher Python
Autonomous hyperparameter search for Sparse Autoencoders on Qwen3.5-0.8B activations. Agent forms hypotheses, runs training cycles, logs to SQLite, and generates its own leaderboard and state summary for the next iteration. Best result: 0.9956 explained variance at 12.5% feature activation.
marl-drone-swarm-rl Python
Multi-agent reinforcement learning for cooperative quadcopter swarm control. MAPPO with centralized critic and shared actor. Curriculum learning from 3 drones on 10×10 grids to 6 drones on 20×20. Physics simulation via PyBullet, PettingZoo multi-agent wrapper.
data-curation-engine Python
LLM-powered pipeline that converts raw documents (PDF, DOCX, HTML, Markdown, source code) into training-ready JSONL datasets. Six-stage pipeline: parse → chunk → generate instructions → rule filter → LLM judge (position-swap bias detection) → output. Multi-provider via LiteLLM. Resumable.
Languages
AI / ML
Backend
Data
Infrastructure
