I am a Vertical LLM & Systems Engineer specializing in building resource-efficient, domain-specific AI models. I bridge the gap between model alignment mathematics and full-stack deployment architecture, replacing bloated, hallucination-prone world-knowledge models with secure, task-aligned intelligence.

The open secret of the industry is that architectural scale is a commoditized trap. For 99% of organizations, prompting a 70B+ generic frontier model is slow, expensive, and a data-privacy nightmare.

My focus is building the alternative: Taking lightweight, open-source base architectures (e.g., Llama, Mistral, Qwen) and executing precise Supervised Fine-Tuning (SFT) and preference alignment (DPO/ORPO) on proprietary vertical data. I build domain experts that outperform frontier models on narrow tasks at a fraction of the inference cost.

Python · Supervised Fine-Tuning (SFT) · Direct Preference Optimization (DPO) · ORPO · Evaluation Harnesses · Benchmarking · Synthetic Data Pipelines · LLM-as-a-Judge · Agentic Workflows · NLP

FastAPI · PyTorch · Hugging Face (Transformers/PEFT) · TypeScript · Node.js · Next.js · Hono.js · PostgreSQL · MongoDB

AWS · Docker · Cloudflare · Nginx · CI/CD Pipelines · GitHub Actions

• 🎯 Domain-Specific Fine-Tuning: Engineering high-fidelity datasets to train highly aligned Small Language Models (SLMs) for regulated industries (Legal, Fintech, Government).
• 🛡️ Sovereign & Air-Gapped Deployments: Leveraging 5+ years of full-stack experience to deploy custom fine-tuned models directly onto secure, on-premise infrastructure.
• 🔬 Rigorous Evaluation: Designing custom domain evaluation harnesses to mathematically track model accuracy and eliminate hallucinations.

I partner with organizations, founders, and venture accelerators looking to build proprietary vertical AI moats rather than renting expensive Western API wrappers.

🌍 LinkedIn: linkedin.com/in/uzairbhayya
✉️ Email: uzairbhayya@gmail.com