Models under constraints
The model is only part of it. The bytes, hardware, and evaluation have to work too.
- Video compression
- Semantic tokens
- Pose estimation
- Integer entropy models
- Tool-use tuning
- GGUF
- MLX
- Quantization
- Evaluation
ML Engineer & a curious person.
Miami, Florida
Data Science & AI · FIU
Independent work

I started in premed. Now I build ML systems in Miami: small models, persistent memory, real hardware. I like work someone else can reproduce.
The questions behind
my work.
The model is only part of it. The bytes, hardware, and evaluation have to work too.
I want to know what happened, what failed, and what evidence a result rests on.
Some of the questions on my desk need hardware, not another browser tab.
Things I have released, repaired, or am still figuring out. Open a folder for the work and its status.
Former #1 · official T4: 0.172
A codec for a deceptively specific problem: keep the information driving models need while making the video tiny. My entry reached #1 on the official T4 evaluation at 0.172. It builds on jas0xf's semantic-token / HPAC work; my contribution includes an integer entropy model, a learned pose basis, and a byte-exact rebuild. Other entrants have since improved on it. That is part of why I published the recipe.
Explore project ↗GGUF / MLX / 4-bit / 16-bit
I fine-tuned gpt-oss-20b for tool use on NousResearch's function-calling dataset using an Unsloth notebook. I published GGUF, MLX, 4-bit, and 16-bit exports so the work can run in different local setups. This is an applied fine-tune, not a claim to have built the base model.
Explore project ↗Hermes / Topcoat / Mnemosyne
The systems I use become the systems I help repair. My upstream work includes a nested-macro formatter panic in Topcoat and conversation-history failures in Hermes Agent and hermes-lcm. The public case studies link the reproductions, fixes, and merge records.
Explore project ↗W3C PROV · ongoing
An unfinished project: an append-only artifact graph using W3C PROV. The idea is to keep the failed attempts alongside the successful ones, with the evidence and lineage still attached. Not released yet. Ask me what is working and what is not.
Explore project ↗What I am building.
How I got here.
Right nowStill in progress
Models, memory, hardware Miami
On my desk: provenance memory, self-healing agents, person search across long videos, and WiFi sensing on four ESP32-S3 nodes. None of these is a finished release.
Ongoing
Hermes / Topcoat / Mnemosyne
When something breaks in the tools I rely on, I want the fix upstream, not just working on my machine.
Making it usable
gpt-oss-20b
One tool-use fine-tune, exported for several ways of running it. The models are public on Hugging Face.
A useful constraint
comma.ai challenge Remote
The comma.ai challenge made the tradeoff concrete: fewer bytes without losing what the driving models need. The result mattered. So did making it reproducible.
The change of direction
Data Science & AI · FIU
I switched because this was the work I loved. It was not the straightest route, but it is the one I wanted to keep following.
Working on a difficult ML problem? Building something in Miami? Or curious about an unfinished experiment? I would like to hear from you.