I build production LLM systems from the metal up — from quantized models running on Jetson edge hardware to multi-agent cloud deployments with tool-use, permission gating, and audit trails. Currently focused on MoE fine-tuning, Blackwell-native FP4 quantization (NVFP4), and agentic coding benchmarks (SWE-bench).
Where my deep work is going right now.
4-bit weights and activations for an 8B MoE on a 16 GB RTX 5070 Ti, custom vLLM CUTLASS SM120 source build. HumanEval 72.6% pass@1 (95% CI [65.3, 78.8]) from a 6.02 GB checkpoint, matching or beating published full-precision results for Qwen 3 7B and Llama 3 8B, and apparently the first HumanEval figure for this model at any precision. 9.5 tok/s single-stream / ~74 tok/s batch-8. Also quantified the reasoning/latency tradeoff: disabling the reasoning trace is ~8.5x faster but costs 17 to 29 accuracy points.
Qwen3.5-4B fine-tuned with ORPO for biblical Q&A — hybrid RAG (ChromaDB + BM25 + cross-encoder reranking), constitutional AI guardrails, and a Whisper + Kokoro voice pipeline behind a Gradio UI. 183 tests, 34 W&B runs.
Autonomous multi-agent personal intelligence on NVIDIA Jetson Orin Nano — 5 LangGraph expert agents, LiteLLM gateway, 3-tier ONNX intent router. Fully on-device, zero cloud dependencies. 415 tests.
Production-grade systems — every one tested, documented, and shipped.
Security-first coding agent. Async ReAct loop with a 4-tier deny-first permission engine, SHA-256 hash-chained audit trail, and 200+ LLM providers via LiteLLM. SWE-bench Lite (dev-23 split): 34.8% single-shot · 52.2% oracle best-of-5, free-tier drivers.
Multi-agent algorithmic trading pipeline with DeepSeek R1 reasoning at every stage. 4-agent pipeline (TA → Chief → Risk → Execution), Kelly Criterion sizing, Monte Carlo risk, real-time WebSocket data.
Comprehensive GPU fleet validation modeled on NVIDIA DCGM. 16 diagnostic modules, Prometheus + Grafana, fault injection, JUnit XML for CI.
ML research control plane — experiment lifecycle management, model registry, and cloud training launcher. Orchestrates preflight checks and knowledge persistence across the stack.
Git-backed knowledge base on Karpathy's LLM Wiki pattern. LangGraph ingest/query pipelines, instructor + Pydantic structured output, BM25 search, Groq → Gemini → Ollama fallback.
SQL + Python ETL pipeline for semiconductor quality analysis — supplier performance scoring, defect Pareto distributions, and yield trend analysis.
Multi-model ML pipeline for Tesla tire wear prediction. Random Forest, XGBoost, and Neural Network ensemble with Claude AI integration.
Upstream contributions on cutting-edge inference hardware.
LLAMA_FTYPE_MOSTLY_NVFP4 case fix (closed, replaced by #22897).From kernels to multi-agent orchestration.