Harsh Nanavati, Applied AI Engineer
I build production LLM systems: RAG, AI agents and evaluation.
I design and ship production AI: retrieval, agents, real-time LLM pipelines and computer vision. At Masterly Solutions I've built six AI systemsEvidence 1 · Selected workSix of the case studies on this page are Masterly Solutions work: three client projects and three in-house products, two of them live. The seventh is a personal project.Go to the work, the largest used by 600+ staffEvidence 2 · DocuPilot – Enterprise RAG PlatformDocuPilot alone is used by 600+ staff in production.Go to DocuPilot's results.
I work remotely from India on IST (UTC+5:30), with 4+ hours of overlap with US East and Europe. Open to full-time, contract or employer of record roles.
Claims on this site link to their evidence. Tap a small number to see it.
Selected work
- DocuPilotEnterprise RAG PlatformAnswers for 600+ staff from company documents, with a source for every answer
- CrickPal Intelligence LayerAI Win-Prediction PlatformAn AI agent writes prediction features, and a test harness decides whether they ship
- CrickPal AI SummaryReal-Time LLM Commentary EngineLive cricket commentary turned into match summaries by a self-hosted LLM
- AI Face Matching SystemGPU face recognition that tags players in tournament photos, unattended
- QuantumCopilotAI Loan & Financial AssistantPlain-English loan questions in, calculated scenarios out, with the LLM never doing the math
- RAG Support & Concierge ChatbotsTwo chatbots, one with RAG and one where I deliberately removed it
- ViperMindMulti-Agent Python TutorFour LLM agents teach Python across 30 topics, while code, not the model, grades every answer
How I build
Answers you can check. If an AI system can't show where an answer came from, people can't trust it. My systems cite their sources, down to the page.Evidence 3 · DocuPilot's two-stage citation gateAnswers cite only passages that clear a minimum score and sit close to the best match, with the exact page region highlighted.Go to the decision
Verify before shipping. Generated code, retrieved context and model output all pass automated checks before they reach a user.Evidence 4 · CrickPal Intelligence Layer's verification harnessGenerated feature code must pass golden-example, mirror-symmetry, replay-leakage and determinism checks, then human approval.Go to the decision
Built to keep running. Retries, job queues, circuit breakers and logging aren't extras. Production AI fails in boring ways, and I design for that from day one.Evidence 5 · AI Face Matching's job queueRetry budgets, expiring leases and parked jobs let it run unattended and recover from bad inputs on its own.Go to the decision
Looking for an Applied AI or GenAI engineering role, remote, full-time or contract. Email me or connect on LinkedIn.