AI/ML ENGINEER · MS @ USF · TAMPA, FL
Harshith
Gujjeti.
▸ |
I build production AI systems, understand the math that makes them work, and ship the ones that move revenue.
SCROLL
I build production AI systems, understand the math that makes them work, and ship the ones that move revenue.
AI/ML engineer and researcher with a sharp focus on production ML systems, generative AI pipelines, and healthcare analytics. I care about building things that actually work in the real world — not just in notebooks.
Currently pursuing my M.S. in AI & Business Analytics at USF, with hands-on research at the Muma College of Business and the Centre for Transportation Research. Previously an AI Research Intern at ISRO.
I build at the intersection of multi-agent AI, computer vision, NLP, and data engineering — wherever hard problems meet real stakes.
Production-ready systems with real data, real metrics, real impact.
Agentic RAG system for structured extraction from pharmaceutical documents — LangGraph orchestration, ChromaDB vector store, RAGAS evaluation pipeline, FastAPI + Docker deployment.
Whisper + spaCy pipeline converting clinician audio into structured SOAP notes. 70% reduction in manual structuring time. Dockerized for HIPAA-compliant local deployment.
Short opinions, held with conviction. Subject to change when the data says so.
If I can't measure the system, I can't improve it. Every LLM pipeline starts with a golden set + RAGAS-style eval loop before I touch the graph. Faithfulness, context precision, answer relevance — numbers first, vibes second.
Fine-tuning is a last resort. 90% of "we need a custom model" is really "we need better retrieval + prompting." I reach for RAG, tool use, and structured output first — cheaper, safer, and swap-in-a-better-model friendly.
A working demo on one small surface beats a roadmap slide forever. Every project I own has a Docker image, a FastAPI endpoint, and one metric I can defend — not just a notebook that runs on my laptop.
"Multi-agent" is glamorous. But every agent loop eventually becomes a deterministic graph with guardrails. I design for that reality — explicit edges, retry budgets, failure modes — instead of hoping the LLM figures it out.
The model is the easy part. Making sure the 150K-row entity resolution actually matches the right companies, or that the 100hr of traffic video gets labeled without drift — that's where the work hides. Own the pipeline end-to-end.
Kubernetes for a weekend project is a red flag. I pick the smallest stack that survives the next 6 months, then spend the saved energy on the part that's actually interesting — the research question, the eval, the user experience.
A year of commits across my open-source work, research pipelines, and side projects.
Training keeps me sharp. The gym teaches progressive overload — a principle I apply just as hard to engineering. Consistency over intensity, every single day.
Gardening grounds me. Watching something you planted grow over time — patience, care, long-term thinking. The same mindset I bring to building ML systems.
I build AI, understand how it works, and apply it where it moves the business. If that sounds useful — let's talk.