Tremayne Timms

ML & AI Engineer — Fine-Tuning · Agentic Systems · Edge Deployment · Production LLM Ops

Dallas-Fort Worth, TX · ttimmsinternational@gmail.com

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).

zaya1-godspeed — NVFP4 W4A4 (4-bit weights and activations) serving for ZAYA1-8B MoE on consumer Blackwell SM120. Custom vLLM CUTLASS source build — 9.5 tok/s single / ~74 tok/s batch-8 (enforce_eager; an earlier 102.6/407 tok/s figure was measured under CUDA graphs, a path confirmed 2026-08-14 to corrupt output on this card and retracted — see repo for the root-cause writeup), 6.02 GB checkpoint on RTX 5070 Ti, plus a validated 2.2× speedup from n-gram speculative decoding on coding-edit workloads, plus s1-style budget-forced reasoning evals proving the checkpoint healthy (GPQA-Diamond 45.8% → 62.5% with reasoning budget).

End-to-end NVFP4 W4A4 quantization + serving for Zyphra's ZAYA1-8B (80-layer MoE + CCA attention) on a 16 GB RTX 5070 Ti. Rebuilt vLLM from source with SM120 CUTLASS FP4 kernels, wrote the layer-wise activation calibration, reverse-engineered the NVFP4 global-scale convention, and built budget-forced eval harnesses for reasoning models. 9.5 tok/s single-stream, ~74 tok/s batch-8 (enforce_eager=True — CUDA graphs corrupt generation on this card regardless of MoE backend, see repo), 6.02 GB checkpoint. Also found and shipped a real batch-1 speedup: n-gram speculative decoding, validated 2.2× on coding-edit prompts, zero training required.

Security-first open-source coding agent. Hand-rolled async ReAct loop with 4-tier deny-first permission engine, SHA-256 hash-chained audit trail, and 200+ LLM providers via LiteLLM. 4,600+ tests.

• 30+ built-in tools with JSON Schema validation, MCP server + client
• Parallel + speculative tool dispatch, cost budget enforcement
• Self-evolution via LLM-guided mutations, multi-language verify gate with retry
• Training data export (openai/chatml/sharegpt), per-step reward annotations for GRPO
• SWE-bench Lite (dev-23 split): 34.8% single-shot · 52.2% oracle best-of-5 — free-tier drivers, $0 API spend; methodology

Autonomous multi-agent personal intelligence system on NVIDIA Jetson Orin Nano Super. 5 LangGraph expert agents, LiteLLM gateway (4 providers + Ollama), 3-tier ONNX intent router. 415 tests. Fully on-device — zero cloud dependencies.

Multi-agent algorithmic trading pipeline with DeepSeek R1 reasoning at every stage. 4-agent pipeline (TA → Chief → Risk → Execution), Kelly Criterion position sizing, Monte Carlo risk simulation, real-time WebSocket market data.

Qwen3.5-4B fine-tuned with ORPO for biblical Q&A. Hybrid RAG (ChromaDB + BM25 + cross-encoder reranking), constitutional AI guardrails, voice pipeline (Whisper + Kokoro TTS), Gradio UI. 183 tests, 34 W&B runs, 5,925 training steps.

• GPU Server Test Suite — GPU fleet validation modeled on NVIDIA DCGM, Prometheus + Grafana dashboards. 188 tests.
• ML Lab — Experiment lifecycle control plane, orchestrates gpu-server-test-suite and llm-wiki. 49 tests.
• LLM Wiki — Git-backed knowledge base, LangGraph ingest/query pipelines, BM25 search. 117 tests.
• ML Experiment Scaffold — GitHub template for reproducible single-GPU ML experiments (config-driven, 3-seed reporting).
• Manufacturing Quality Analytics — SQL + Python ETL for semiconductor quality analysis.
• Tesla Tire Wear ML — Multi-model ensemble for tire wear prediction.

• llama.cpp #22897 — NVFP4 default type mapping + per-tensor scale tensors + MSE-optimal correction (open, awaiting upstream review)
• llama.cpp #22858 — Missing LLAMA_FTYPE_MOSTLY_NVFP4 case fix (closed, replaced by #22897)

📈 Contribution Graph