Yash Watwani

Bengaluru · [email protected] · linkedin.com/in/yashwatwani · github.com/yashwatwani · yashwatwani.com

Summary

Lead Data Scientist and AI engineer with 8 years building production ML, deep learning and generative AI systems across marketing, banking, insurance and restaurant tech. At Pyvot I build LLM products end to end: agent design, evaluation, cost optimisation, privacy and AWS deployment. Before that at Brainlabs I built an agentic marketing-mix-modelling platform for Google, Bayesian MMMs and forecasting, all shipped to production.

Experience

• Pia, Mynt's AI analyst (owner, end to end). Restaurant owners ask about payouts in 5 languages and get real figures with an auto-chosen chart. LangGraph on FastAPI + PostgreSQL (AWS): triage, a self-correcting tool-calling SQL writer, and an embedding shortcut over ~1,300 common questions.
• Hallucination control. Every number, including derived % changes and totals, is verified against the query results: unverified answers 46% → 4%. Nightly grading on 109 real questions, ~2,960 automated tests, a 324-question model-comparison bank.
• Cost and reliability. ~95% cheaper per step on Claude Haiku, one-hour prompt caching, one interface over 6 LLM vendors with per-step failover, OpenTelemetry tracing.
• Security. Tenant isolation inside PostgreSQL, a SQL allowlist parser, and a prompt-injection screen that caught 85/85 attacks.
• Ethos Menu AI. Reads a restaurant's menu (PDF, photo or sheet), re-engineers it against a 90-rule rulebook and returns upload-ready Zomato and Petpooja files.
• Product analytics. First-party, privacy-safe event tracking for the website, dashboard and app on AWS ECS + Aurora + Metabase.

• Agentic MMM platform for Google. Multi-agent system on Google ADK + Vertex AI that runs marketing-mix workflows end to end on Cloud Run.
• Bayesian marketing mix models in PyMC and Meridian for FMCG and beauty brands: channel ROI with uncertainty, response curves, budget scenarios.
• Enrolment forecasting on Databricks with LightGBM/CatBoost ensembles, beating budget baselines on WMAPE.
• Automated weekly reporting from 7 BigQuery tables to KPI PDFs across 4 markets, replacing a manual process.

• LSTM email classifiers for banking and insurance, and a German-language extraction pipeline.
• Document intelligence: blur detection, document clustering, spaCy + AWS Textract key-value extraction.

• Signature detection and verification with Mask-RCNN and Siamese networks for bank fraud teams.
• A COVID-19 research search engine over 77,000+ papers with BERT and USE semantic search.
• County-level US surgical-supply demand forecasting during the pandemic peak.

Skills

GenAI / LLMs

Agents, text-to-SQL, tool calling, RAG, LLM evals, prompt caching, prompt-injection defence

LLM frameworks

LangGraph, LangChain, Google ADK, LlamaIndex, Arize Phoenix, LangSmith

ML / AI

Deep learning, NLP, computer vision (YOLO, Mask-RCNN, SAM), forecasting, marketing mix modelling

Statistics

Bayesian modelling, regression, clustering, cohort analysis, backtesting

ML frameworks

PyTorch, TensorFlow, Keras, scikit-learn, LightGBM, CatBoost, XGBoost, PyMC, Meridian, Hugging Face

Code / APIs

Python, SQL, TypeScript, FastAPI, Flask, Node.js

Data

PostgreSQL, BigQuery, Metabase, Pinecone, ChromaDB, FAISS, Milvus

Cloud / MLOps

AWS, GCP, Databricks, Azure, Docker, GitHub Actions, MLflow, OpenTelemetry

Education