TaSain Valdez Thomas
AI Engineer who translates published ML research into working systems — calibrated BMI estimation with conformal prediction, LLM-as-judge reasoning, and a self-hosted homelab for private AI inference.
What I've built
Research-to-prototype systems at the intersection of AI and infrastructure.
ConformalBMI — Uncertainty-Aware Body Mass Estimation
Translated published BMI-estimation research (Digital Scale, Sui et al., Manichand et al., Jin et al.) into a working estimator — a Squeeze-and-Excitation DenseNet (SE-DenseNet121) in PyTorch, extended with split conformal prediction for calibrated uncertainty and a GPT-4.1-Nano reasoning layer that explains each estimate.
- Research-to-prototype: built on peer-reviewed BMI estimation literature, extended with conformal prediction and LLM-based reasoning
- Split conformal prediction for distribution-free uncertainty — 90% intervals with empirically verified 90.4% coverage
- 2.81 BMI-point MAE (11.56% MAPE) on Celeb-FBI, a 7,208-image real-world benchmark
- GPT-4.1-Nano LLM-as-judge explains each estimate and flags edge cases in plain English
Research Foundation
Built on established BMI estimation research and extended with split conformal prediction for calibrated uncertainty and GPT-as-Judge reasoning:
Homelab — Privacy-First AI Inference
Self-hosted homelab for privacy-first AI serving — a Proxmox hypervisor with GPU passthrough running an AI agent inside a dedicated virtual machine.
- Configured memory and resource allocation across nodes; assembled and connected the hardware end to end
- Built for data privacy — serving models locally without third-party cloud dependencies
On-Wheels Detailing — Founder & AI Agent Orchestration
Founded and operate a detailing business with paying customers; wired up Stripe payment processing end to end.
- Orchestrated a self-hosted AI agent to design and generate the business site via prompt-engineered role-play, with a no-customer-data-persisted privacy constraint
- Built a Telegram agent layer delivering daily morning briefings and real-time lead alerts
- Automated scheduled social content curation (M/W/F) with hashtag and engagement targeting
- Deployed on DigitalOcean with Nginx, systemd, and Cloudflare DNS/SSL
Client Project — PDF Editing & File Management (Flutter)
Capstone project — Wayne State University. Team lead of a 4-person team building a mobile collaboration app for a real client, layered on an existing product (flow charts, Venn diagrams, accounting tools).
- Personally built the PDF file-management feature — an annotation layer that overlays text on PDFs and lets changes be saved, edited, or deleted
- Ran the team process end to end — three weekly meetings, client communication, and deadline adherence
- Proposed the GitHub workflow — branching, pull requests, and issue tracking
Where I've worked
Built a worker-to-station assignment engine (C++/SQLite) adopted by the plant and used during real shift changes. Developed an AR-based vision-inspection system (Vuforia Model Targets + Inductive Automation Ignition). Built managerial monitoring systems with role-based authorization and audited write access.
Founded and operate a detailing business with paying customers. Wired up Stripe end to end and orchestrated a self-hosted AI agent that designs the site, delivers daily briefings, and alerts on real-time leads.
Academic background
B.S. Computer Science
Trustworthy AI coursework: LoRA/QLoRA, RLHF/DPO/GRPO, Red Teaming, Machine Unlearning (TOFU/WMDP), Agentic AI Safety, Grounded VLMs (LISA/GLaMM), Chain-of-Thought Reasoning, HarmBench/SafeChain.
Technologies I work with
Hands-On
Coursework
Get in touch
Available for opportunities in Detroit, Michigan (EST).
Detroit, Michigan · EST (UTC-5)
Onsite preferred
Work history
A track record across AI, automation, and full-stack engineering.
Industry 4.0 / Automation Intern (Rotational)
Worker-to-station assignment engine adopted by the plant; AR vision-inspection system; monitoring dashboards with role-based access.
Founder & AI Agent Orchestration
Real detailing business with paying customers. Stripe payments, a self-hosted AI agent, and a Telegram layer for briefings and lead alerts.
CS Graduate — Trustworthy AI
GPA 3.4. Specialized in AI safety: LoRA, RLHF, red teaming, machine unlearning, grounded VLMs.