Headline

AI Engineer / Staff Data Scientist | Agentic Systems (LangGraph, Multi-Agent Orchestration) | Production RAG & Hybrid Retrieval | Causal Inference, Experimentation & A/B Testing | LLM Evaluation

About

Staff Data Scientist & AI Tech Lead | Agentic AI · Production RAG · Causal Inference · Experimentation Platforms 14 years turning ambiguous business problems into production AI systems — and owning the result after they ship. ▪️Agentic AI & LLM Platforms Set the architecture for a multi-agent data modelling platform at ServiceNow — LangGraph, MCP, master/domain agents, human-in-the-loop review gates, golden-set eval harness in CI/CD. Enterprise data model design: 2–3 weeks → under 2 hours. ▪️Production RAG at Scale Built the support intelligence platform running at 60k+ interactions — hybrid BM25 + dense retrieval with RRF fusion, cross-encoder reranking, semantic chunking, tenant-level ACLs, abstention thresholds, semantic caching. Context discovery: hours → minutes. ▪️ LLM Evaluation & LLMOps Retrieval (Recall@k, nDCG) and generation (faithfulness, context precision) scored separately behind human-calibrated LLM-as-a-Judge gates. Built the eval framework other teams now use as the reference pattern. ▪️ Causal Inference & Experimentation Architected the global A/B testing platform at AB InBev — 200+ live experiments/month across NA, MAZ and APAC. CUPED variance reduction cut sample sizes 30–40%. Operationalised Generalised Synthetic Control for high-noise store-level tests; inconclusive outcomes fell from ~65% to ~20%. Work behind the 2022 INFORMS Daniel H. Wagner Prize (ABI–MIT). ▪️ Commercial Impact $2B+ raw-material spend base · $50M+ enterprise renewal portfolio · 30% digital acquisition-cost reduction for a global QSR client — each with an explicit measurement design behind it. Tech: LangGraph · LangChain · MCP · RAG · pgvector · FastAPI · Kubernetes · Python · SQL · PySpark · Snowflake · Databricks · XGBoost · PyTorch · AWS / Azure / GCP Methods: Multi-Agent Orchestration · Tool Calling · ReAct · Hybrid Retrieval · Cross-Encoder Reranking · LLM-as-a-Judge · CUPED · Generalised Synthetic Control · Difference-in-Differences · Double ML · Uplift / CATE Modelling · Time-Series Forecasting Open to senior IC and technical leadership roles — remote or relocation.

Location

Greater Bengaluru Area

AI stack

anthropic apis

claude code

claude

gemini

langgraph

langsmith

Work experience

Staff Data Scientist / Tech Lead at ServiceNow, May 2024 - Present · 2 yrs 4 mos, Senior IC setting technical direction for agentic AI and causal measurement on the enterprise data platform. ▪️ AGENTIC DATA MESH — MULTI-AGENT DATA MODELLING Led design and build of a multi-agent platform (LangGraph, MCP, FastAPI, Snowflake, Kubernetes) — master agent directing domain agents to generate Kimball dimensional models, fact tables and ETL DDLs Resolved names against the data catalog and knowledge graph first, calling the LLM only for ambiguous cases — kept cost and output predictable Four human-in-the-loop review gates + golden-set CI/CD eval harness on trajectory and tool-call accuracy Model design turnaround: 2–3 weeks → under 2 hours ▪️ SUPPORT INTELLIGENCE — PRODUCTION RAG, 60k+ INTERACTIONS Built production RAG for grounded response synthesis and ticket routing Hybrid BM25 + dense retrieval fused with RRF, cross-encoder reranking, semantic chunking, tenant-level ACLs, abstention thresholds, semantic caching Retrieval (Recall@k, nDCG) and generation (faithfulness, context precision) scored separately behind human-calibrated LLM-as-a-Judge gates in CI/CD Context discovery and ticket resolution: hours → minutes ▪️ CAUSAL MEASUREMENT OF AI FEATURES Leadership was justifying AI investment on adoption-vs-retention correlation — but adopting accounts were already the healthiest Built the causal framework (covariate matching, confounder controls, treatment-effect estimation) to isolate true incremental retention lift Agreed the evidence standard with Product and Customer Success before results existed Defensible CXO read on AI impact across a $50M+ renewal portfolio ▪️ HIRING & TECHNICAL LEADERSHIP Own technical and system-design rounds on the DS/AI hiring panel; wrote the bar for agentic and causal candidates Internal readouts made the team's review and eval practice the pattern other groups adopted

Data Science Manager at AB InBev, Sep 2022 - Jul 2024 · 1 yr 11 mos, • Led cross-functional team to build AB InBev’s company-wide A/B testing platform from zero — scaled to 100–200 concurrent experiments/month; introduced CUPED variance reduction cutting required sample sizes by 30–40%; established A/A testing, pre-experiment randomisation diagnostics, and standardised MDE frameworks. → Impact: Reduced experiment runtime by ~30% for seasonal decisions; self-serve dashboards cut DS dependency and enabled business teams to run experiments independently. • Co-developed and published Generalised Synthetic Control (GSC) model with MIT (Wagner & Ward framework) — GSC applies low-rank matrix factorisation to the store × week sales panel to estimate the no-treatment counterfactual via PCR and recover an unbiased ATE; runs on every A/B measurement cycle. → Impact: Replaced correlation-based reporting with auditable causal lift estimates; directly powers promotional spend allocation across multi-crore annual budgets. • Architected Multi-Horizon Demand Forecasting and Nowcasting Platform — AB InBev must commit to raw material contracts (barley 9 months out, cans 3 months out, production 2 weeks out) simultaneously; three-model reconciliation : SARIMAX (H1: 0–2 weeks, 900K daily SKU models), Prophet with event calendar + weather regressors (H2: 1–3 months), Kalman Filter + ARDL (H3: 3–9 months); Kalman nowcasting fills sell-out data lag; MinTrace ensures cross-horizon coherence. → Impact: Enabled calibrated procurement commitments across a $2B+ annual raw material spend base; reduced overstock write-offs by improving demand signal accuracy at each planning horizon. • Awards: Top Performer (2023), Stellar Pitcher and Pint Award (2023)

Senior Data Scientist-II / Lead MLE at Publicis Sapient, Apr 2020 - Aug 2022 · 2 yrs 5 mos, On-site, • Designed uplift modelling framework for a global QSR chain’s digital app campaigns — identified persuadable segments (top 20–30% of users), preserving full campaign effectiveness through precision targeting across multiple markets. → Impact: 30% reduction in digital acquisition costs. • Built two-tower deep learning recommendation system on AWS SageMaker (user + item embedding towers on behavioural signals); continuous A/B validation post-deployment. → Impact: 2% lift in Average Order Value.

Data Scientist at Landmark Group, Sep 2018 - Apr 2020 · 1 yr 8 mos, Greater Bengaluru Area, Built and deployed production ML systems for one of the Middle East’s largest retail conglomerates. Demand Forecasting — Ensemble forecasting system combining ARIMA, XGBoost, and LSTM. Deployed as APIs on AWS SageMaker with automated monitoring and retraining triggers. Reduced average inventory costs by 15%. Churn Prediction — Built customer churn models using ensemble methods (LightGBM + XGBoost + Random Forest). Achieved 10% reduction in lapsed customer churn rate.

Data Science Consultant at Lenovo, Aug 2017 - Aug 2018 · 1 yr 1 mo, Bengaluru, Karnataka, India, Developed & deployed Employee Attrition Modeling, build sophisticated text processing pipeline for extracting appraisal, survey data using NLP techniques like (Word Embedding, Avg-Word2Vec, Tf-IDF) tied with headcount & incentive data and build powerful Machine learning model using XgBoost, CatBoost, LightGBM and other algos • Design a 3 *3 Design Risk vs. Performance dashboard to strategically analyze, solution impacted Employee retention by 3-5% by target them at right time

Education

Master of Business Administration - MBA, Management Information Systems, General, Maulana Abul Kalam Azad University of Technology, West Bengal formerly WBUT

Bachelor of Science - BS, Economics & Statistics, Ramkrishna Mission residential college,Narendrapur