Harsh Jha

I’m a Founder & AI Systems Developer at DoubtAI, where I design and build AI systems focused on measurable reliability. Since 2025, I’ve worked on RAG that doesn’t just retrieve content, but also evaluates retrieval and output quality with repeatable logic.

I built a custom RAG evaluation loop by embedding 3,032 vectors from 10 textbooks, measuring retrieval quality via cosine similarity, and iterating chunk size and overlap to maximize recall. I also implemented deterministic multi-model routing logic with structured prompt templates to keep LLM output consistent and machine-parseable.

To reduce waste without losing correctness, I added MD5 hash-based answer caching validated against expected outputs, cutting redundant API calls by ~40% while verifying response consistency. I’m especially driven by assertion-based grading (exact match, substring checks, similarity thresholds) and leakage-avoidance methods, and I enjoy shipping these systems end-to-end with Python, FastAPI, and CLI tooling.