Bhargav Raghavendra
Summary
I take an undocumented problem, own it from the first conversation to production, and ship the solution into the customer's own environment. I work across the whole stack, from LLM agents and RAG to full-stack and infrastructure, and I am at my best on the part nobody measures: I find it, build the instrument that measures it, then follow what it says even when it kills the plan. I find the bottleneck, build the thing, and put it in front of someone.
Experience
- Audited 8,831 generated questions and found the difficulty labels were an artifact: on hard questions the correct answer was the longest option 85% of the time, beatable without reading the question. Rebuilt generation to remove it.
- Took it to production: a RAG content pipeline over ChromaDB with BM25 re-ranking at $0.30 to $1.50 of model and speech spend per concept, six containerised services with CI on every repository, and 138 versioned migrations that abort a deploy on failure so the last good build keeps serving.
- Built the school go-to-market myself, from ICP and target lists through outreach to the offer I put in front of schools, then read the demand signal honestly and repositioned onto the teacher workflow the research surfaced.
- Worked directly with UCLH clinicians to validate real surgical procedures, turned what they demonstrated into interaction flows that held up in the headset, and shipped the experience for hospital use as an early team member.
- Built and evaluated machine learning models in Python to classify water quality from real-world datasets, owning cleaning, features, training, validation and error analysis.
Projects
Answers what a controls team asks before go-live, whether an agent can be trusted to approve invoices. Caught an agent that reached the correct outcome on 191 of 192 runs while still breaking a mandatory control on 7 that completed, which no output-level eval can see. Declares the invariants a run must never violate, perturbs the input in ways a correct agent should ignore, and localizes each breach to a named step with a Wilson confidence interval. Found my own headline result rested on an unfair baseline and published the correction instead. 2,082 runs committed for recheck.
Moves control flow off the language model into a six-state machine with a static transition table, so the same document yields a byte-identical session plan across 40 replay runs, compared by SHA-256. A resume index returns a learner to the exact unit they left. Basis of patent IN202641086881.
Reads a team's own Postgres and Supabase migrations and answers the question that blocks a go-live: can one tenant see another tenant's rows. Returns a counterexample with the SQL that reproduces it, or a proof that no such leak exists, never a guess. I wrote the SQL parser and a from-scratch decision procedure that handles NULL correctly under three-valued semantics, then differential-tested it against Z3 on 3,200 generated formulas.
Reads live Greenhouse, Lever and Ashby boards for a company universe I curated by hand and flags which accounts have a buyer-persona role open right now, every row linked to the exact posting. The thesis lives in JSON, so buyer titles, exclusions, weights and geography retarget to a new sector without touching code. Maps 81 companies across robotics and B2B SaaS operations.
Patents
Skills
Leadership
- Ran the organisation as Secretary across 20+ community programs, owning governance, approvals and university clearances, and built partner-NGO outreach that placed 500+ student volunteers.