TaSain Valdez

Secure AI Pipeline for BMI Estimation

Fine-tuned a 7B Vision-Language Model via LoRA, compressing trainable parameters to 35M (0.5%) while achieving 89% accuracy with 95% lower inference costs.

• Privacy by design: face blurring, synthetic data (BEDLAM), and LLM-as-Judge evaluations
• Reduced hallucination & refusal rates by 40% via custom refusal mechanisms
• Chain-of-Thought reasoning with zero-trust input validation, exceeding HarmBench safety benchmarks

MLOps & AI Agent Infrastructure

Productionized an auto-scaling Kubernetes inference cluster with GPU passthrough on Proxmox, achieving 99.9% uptime for LLM serving.

• Embedded red-team testing (Kali Linux) into CI/CD — patching prompt injection vulnerabilities pre-production
• Automated lead alerts via cron-driven scripts, reducing sales response time from hours to real-time

SaaS Booking Platform — On-Wheels Detailing CRM

Production CRM supporting multi-location, 6-tier pricing with Stripe payments and Telegram bot integration.

• Deployed on DigitalOcean with Nginx, systemd, and Cloudflare DNS/SSL/CI/CD
• 24/7 uptime with auto-restart policies — handled 100% of production incidents post-launch

Inventory Management Website — Detroit Business

Engineered a custom inventory management web application for a local Detroit business, handling product tracking and order management.

Magna Automation — Production Line Assignment Engine

Architected a worker-to-station matching system that reduced manual assignment time by ~40% during shift changes at a major automotive manufacturer.

• Built evaluateFit() scoring algorithm balancing line familiarity, station expertise, and employee rank
• Database-driven configuration enabled runtime reassignment without code changes