Harsh Nanavati

Applied AI Engineer: LLMs, RAG, agentic systems and evaluation

nanavatiharsh00@gmail.comLinkedInAhmedabad, India (IST), open to remote roles

Download PDF

Summary

Applied AI engineer with 2+ years shipping production LLM systems end to end: enterprise RAG, agentic pipelines with automated verification, real-time self-hosted LLM services and GPU computer vision, for clients in the US and India. Systems serve 600+ users and cut manual work by up to ~70%. I own the full path: retrieval design, guardrails, evaluation, FastAPI backends and deployment.

Experience

SDE – Artificial Intelligence (Applied AI Engineer)

Oct 2024 – present

Masterly Solutions Pvt. Ltd., Ahmedabad, India

Building in-house AI products and client systems for the US and India.

  • DocuPilot – Enterprise RAG PlatformCut manual SOP queries by ~40% for 600+ staff across 14 document categories, with Docling ingestion (OCR and tables), custom structure-aware chunking with page-level provenance, and text-embedding-3-small with FAISS retrieval over 12,000+ chunks at sub-100 ms median latency.
  • DocuPilot – Answer Grounding & GuardrailsGPT-4o-mini answers constrained by answer-grounded citation selection and a two-stage citation gate, with every claim linked to its source page region; role-based document access, JWT auth, SHA-256 dedup and per-query observability on FastAPI and SQL Server.
  • CrickPal Intelligence Layer – AI Win-Prediction PlatformAn LLM agent turns analysts’ plain-English specs into feature code, gated by an automated verification harness (golden examples, symmetry, replay-based leakage, determinism) with self-repair and human approval; Elo baseline at log-loss 0.643 and Brier 0.227 over 4.1k T20Is; ~800 automated tests.
  • CrickPal AI Summary – Real-Time LLM Commentary EngineLive ball-by-ball data turned into narrative and headline summaries on self-hosted Ollama; incremental summarization keeps prompt size flat as a match grows; exactly-once batching on a single instance, a circuit breaker with backoff, and an admin console for per-match prompt and model tuning without redeploys.
  • AI Face Matching SystemCut manual photo tagging by ~70% with an unattended MTCNN and FaceNet pipeline: self-healing job queue, retry budgets, stale-lease recovery and an embedding cache across three SQL Server databases; extended to video with YOLOv8 detection and ONNX/DirectML export.
  • QuantumCopilot – AI Loan & Financial AssistantMaps natural-language queries to typed parameters across 6 loan products; parallel scenario fan-out (~3x faster scenario stage), multi-turn memory, and 8 Uvicorn workers on IIS.
  • RAG Support & Concierge ChatbotsA support bot with a human-in-the-loop correction loop (~60% fewer escalations); a concierge bot redesigned from RAG to cached long-context with single-call intent routing.

Intern – Data Science & Big Data

Jun 2023 – May 2024

Deloitte Haskins & Sells LLP, Remote, Ahmedabad, India

  • Used Spark, Scala and SQL for entity-resolution and anomaly-detection analyses on large-scale financial datasets.
  • Built Power BI dashboards for enterprise audit reporting workflows.