Apoorva
Verma
AI Engineer building retrieval systems that don't hallucinate — and the evaluation harnesses that prove it.
This site is one of them. Every answer below is retrieved, grounded, and scored before it reaches you.
How this site's AI is evaluated
Every answer above is scored before you see it. I ran an ablation across chunking strategies and embedding models — the winning configuration is what's serving you now.
Experience
AI/ML Engineer · Oracle
2023 — nowRAG systems & LLM evaluation infrastructure for enterprise search.
SWE Intern · Microsoft Engage
2022Built an ML-driven recommendation feature end-to-end.
SWE Intern · J.P. Morgan
2021Data tooling for an internal risk-analytics platform.
Fellow · MLH
2021Open-source fellowship — shipped to a production codebase.
Featured projects
RAG + Evaluation System
Production RAG pipeline with semantic chunking, hybrid retrieval, and an automated RAGAS eval gate blocking regressions in CI.
LLM Eval Benchmark
Open benchmark harness comparing 8 LLMs across faithfulness, toxicity, and instruction-following with LLM-as-judge calibration.
AI Agent — failure recovery
Tool-using agent with self-critique loops and graceful degradation — detects failed tool calls and re-plans instead of hallucinating.
Notes on grounding & eval
View all writing →Engineer who measures what others guess
I'm an AI engineer focused on the unglamorous half of LLM work: making sure systems are actually correct. At Oracle I build RAG pipelines and the evaluation infrastructure that keeps them honest in production.
Before that, a path through some places that taught me to ship: Microsoft Engage, J.P. Morgan, and Major League Hacking. I care about reproducible benchmarks, calibrated judges, and answers that cite their sources.