Remote roles
Remote machine learning engineering, full-time or contract. I am US-based in Central time. Email me the role and the stack and I will reply with a resume.
About
Donald Galliano III
Machine learning engineer with a background in healthcare.
where: New Orleans, Louisianastatus: taking work
- RHIA
- MHA
- RPSGT
- M.S. CS
- AWS MLE
I'm Donald Galliano III, a machine learning engineer in the New Orleans area, currently at CRC Global Solutions. Before that I spent years inside hospital sleep labs and health information management, which is where I learned what clinical data looks like before anyone cleans it.
How I work
Two ways in, one address.
Local contracting through Crawdata
In the New Orleans area I take scoped builds, performance engineering, and custom models on contract. Fixed scope or retainer, on site when it helps. Send me the problem: what it is, where it runs, and what it costs you.
Either way: hello@crawdata.ai. I reply within two business days.
Where I have worked
Clinical floor, classroom, production ML.
CRC Global Solutions
Machine Learning Engineer, full time
Current. Developing AI-native brokerage and transportation systems.
LCMC Health
Registered polysomnographic technologist
Overnight sleep studies in a hospital lab. Epic, HL7 and FHIR, and HIPAA were the daily environment.
Delgado Community College
Faculty, sleep technology program, and advisory board member
I teach the AI side of sleep medicine: the algorithms inside home sleep apnea testing and auto-titrating PAP, in depth, so students know what the device is deciding and why.
Work
Selected highlights.
Clinical deterioration classifier
88.26%
Deterioration-severity classifier over emergency department vitals and raw ECG waveforms across 2.1M rows. A gated three-stage ensemble (Gaussian process, Bayesian neural net, BiLSTM) with 95.5% recall on the severe class, where a miss costs the most.
30B fine-tune on a shoestring
30B / $300
LoRA fine-tuned a 30-billion-parameter Mamba-Transformer MoE on about $300 of cloud, fed by roughly 670K programmatically verified synthetic reasoning examples I generated myself.
djinnax
240×
Batch-native JAX simulation engines, parity-gated against the reference before any timing, benchmarked on a consumer GPU. Open source, code and numbers public.
Food Bank Visit Bot
5 at once
Pro bono desktop app for two food pantries. It logs visits into Feeding America’s Neighbor Intake site so volunteers stop retyping spreadsheets, running up to five visits in parallel: 26 seconds for five versus 38 for one. Electron, with a Cloudflare Worker for shared rosters.
Claude Token Tracker
1 file
A usage dashboard for Claude Code in one 687-line Python script with no dependencies. It reads the transcripts already on disk, including every subagent, and renders a static page of cost per day, per model, per project, and per agent. No JavaScript in the output, charts as inline SVG, built to be read without color. MIT.
Greener Pastures
369 tests
A Minecraft mod that treats a Cobblemon breeding line as a data pipeline: a real web console rendered in-game, node-graph filters, and a server-authoritative arcade with unit-test-pinned odds. MIT, on CurseForge.
Featured
Teaching a policy to play an eight-player auto-battler
- 80%top-4 finishes, first-gen agent, 50 games
- 812actions, current schema
- 46,196observation floats, current schema
- 88Mpolicy parameters, current schema
Every reward signal in this project is documented with its cadence and magnitude, so reward hacking can be found by reading, not guessing. That habit is the point of the project. The game is Pokemon Auto Chess, an open-source, eight-player auto-battler with a 1,139-species roster, 31 synergies, items that combine, and an economy.
I forked the 42,000-line TypeScript game and built a headless training server on top of it. Transport started as HTTP, then moved to a binary Unix-socket protocol and a shared-memory vector environment so PPO could roll out many games at once. The agent trains with MaskablePPO on a curriculum: one seat against seven scripted bots, then a hybrid with several agents in the same game, then full eight-seat self-play. The first-generation agent, on a 612-float observation and 92 actions, finished top four in 80% of a 50-game evaluation. The current schema is a 46,196-float observation and 812 actions, including manual item equips, driving an 88M-parameter policy that has not been evaluated yet.
The simulation core is being ported to Rust with libm pinned so combat math matches the original bit for bit, checked by parity fixtures. A Chrome extension captures live games straight from the browser, 352 human games to date, analyzed for economy curves, positioning, and streak effects. The same extension ships publicly as a real-time roll-odds calculator.
the extension on the Chrome Web Store · the numbers on the home page
Background
How I got here.
In high school I tested into a robotics program run by the US Marine Corps and Air Force, and between Asimov and embedded systems I settled on one goal early: work where humans and technology meet.
For a while that meant healthcare. Clinical bedside work, health information management, and healthcare analytics taught me what real data looks like before anyone cleans it. Then I found machine learning and haven't put it down since. Healthcare gave me the problems. ML is where I want to solve them.
Most of what I build is above: clinical models, competition entries, tools for people who needed them, and game AI. The common thread is a model that has to survive contact with the real thing.
Find me
Elsewhere.
Donald Galliano III: Registered Health Information Administrator (RHIA), Master of Health Administration (MHA), Registered Polysomnographic Technologist (RPSGT), Master of Science, Computer Science (Machine Learning) (M.S. CS), AWS Certified Machine Learning Engineer (AWS MLE). Founder, Crawdata.ai.
Contact
Hiring, or have a problem?
Recruiters: send the role and the stack and I will reply with a resume. Clients: send the problem, where it runs, and what it costs you. You work with me directly either way.
hello@crawdata.aiFrom the mud to models