Skip to content
View t-timms's full-sized avatar
💭
God is the source code
💭
God is the source code
4 followers · 0 following

Achievements

Achievement: Pull Sharkx3Achievement: YOLOAchievement: Pair Extraordinairex3Achievement: Quickdraw

Achievements

Achievement: Pull Sharkx3Achievement: YOLOAchievement: Pair Extraordinairex3Achievement: Quickdraw

Block or report t-timms

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
t-timms/README.md

Tremayne Timms

ML Engineer — LLM quantization & inference on consumer hardware

Dallas–Fort Worth, TX · Portfolio · Hugging Face · LinkedIn · ttimmsinternational@gmail.com

Python, PyTorch, Docker, GitHub Actions, TypeScript, PostgreSQL

I quantize and serve large models on hardware that isn't supposed to run them — 16 GB Blackwell GPUs, Jetson edge boards — and I reproduce every published number against its confidence interval before I call it done.

Open to ML Engineer roles — DFW or remote.

Projects

MoE Pruning + NVFP4 — A 50%-expert-pruned MoE coder model, quantized to fit 16 GB VRAM. SWE-bench Verified 52.0% (26/50, officially graded), HumanEval+/MBPP+ reproduced inside published confidence intervals, CI-checked reproduction pipeline. Model on Hugging Face →

ZAYA1 NVFP4 W4A4 — 4-bit weights and activations on native Blackwell tensor cores: 9.5 tok/s single-stream from a 6.02 GB checkpoint, 2,100+ combined downloads on Hugging Face. Includes a benchmark I retracted and corrected in public once I found the CUDA-graph path corrupting output. Model on Hugging Face →

Godspeed Coding Agent — A coding agent built from scratch: deny-first permission engine, SHA-256 hash-chained audit trail. SWE-bench Lite 34.8% single-shot / 52.2% oracle best-of-5, $0 API spend.

Sovereign Edge — Five-agent personal AI system running entirely on a Jetson Orin Nano — zero cloud dependencies.

Also: Bible AI Assistant (ORPO fine-tune + hybrid RAG) · Manna Trading (multi-agent trading pipeline) · an open llama.cpp PR fixing an NVFP4 quantizer crash

Stack

Python · PyTorch · vLLM / CUTLASS · TRL / Unsloth · NVFP4 · GGUF · GPTQ / AWQ · CUDA · Docker

Pinned Loading

  1. bible-ai-assistant bible-ai-assistant Public

    Domain Q&A assistant - ORPO fine-tune, hybrid RAG, guardrails, voice pipeline.

    Python

  2. kat-coder-nvfp4 kat-coder-nvfp4 Public

    Serving a large MoE coder model on consumer Blackwell GPUs via REAP expert pruning + NVFP4 quantization.

    Python

  3. zaya1-nvfp4-w4a4 zaya1-nvfp4-w4a4 Public

    NVFP4 W4A4 quantization + serving for ZAYA1-8B MoE on a 16 GB RTX 5070 Ti, with reproducible throughput benchmarks.

    Python 1

  4. godspeed-coding-agent godspeed-coding-agent Public

    Security-first open-source coding agent - deny-first permission engine, hash-chained audit trail, SWE-bench Lite evaluated.

    Python

  5. sovereign-edge sovereign-edge Public

    Local-first personal AI stack on Jetson Orin Nano Super.

    Python

  6. llama.cpp-nvfp4 llama.cpp-nvfp4 Public

    Forked from ggml-org/llama.cpp

    llama.cpp fork carrying my open NVFP4 quantizer patch - upstream PR ggml-org/llama.cpp#22897.

    C++