Production ML · Agentic AI · LLM Evaluation

Data Scientist

Production ML, agentic AI, and LLM evaluation. Eight years shipping systems in regulated industries.

~8 years building data systems across pharma compliance at Amgen and financial services risk at American Express. Both regulated industries. Both demand interpretability and rigorous evaluation.

Three end-to-end portfolio projects — a clinical-trial RAG system with agentic Corrective-RAG and a cross-encoder reranker, a patient risk stratification pipeline with causal inference and survival analysis deployed live on AWS Lambda, and a pharma compliance spend analytics system on 16M+ CMS Open Payments records across Pandas, Snowflake, and PySpark.

View Projects Download Resume

Years across
pharma & finance

End-to-end
projects live

105+

Markets & products
risk-scored

16M+

CMS records
analyzed

About

Eight years inside regulated industries.

I'm a Data Scientist with ~8 years of work across pharma and financial services. At Amgen (2022–2026), I architected the Python ETL behind the company's CMS Open Payments federal submission (~509K records, ~$630M in the latest cycle), built a multivariate risk scoring system across 65+ ex-US countries and 40+ US products presented to Directors and Executive Directors, and led an NLP initiative that cut manual document categorization by 25%. At American Express via Atos Syntel (2016–2019), I built risk scoring pipelines on 10M+ transactions and the team's A/B testing framework that ran 12+ cycles per year for two years.

My work has always lived inside regulated industries where explainability matters more than fluency. A model that can't be defended to a regulator doesn't ship. A score that can't be traced to a feature doesn't get used. Those constraints are why I lean into causal inference, rigorous evaluation, and audit-defensible methodology — not the slowest path, just the only one that survives contact with production in healthcare and finance. I build with AI tooling daily — Claude Code and Model Context Protocol integrations are part of my working setup — and I review everything they produce against the same bar.

I'm targeting Senior Data Scientist, ML Engineer, and AI Engineer roles where production ML, agentic AI, and rigorous evaluation matter. Open to remote, hybrid, or onsite for the right opportunity. US citizen — no sponsorship required. If your team works on hard problems in regulated industries — or just wants to ship ML that holds up — I'd love to talk.

How I Work

Four principles in every project I've shipped.

Principle 01

Test on your domain before you trust the benchmark.

The biggest unlock in my Clinical Trial RAG project was switching from MiniLM to PubMedBERT — not because the leaderboards said so, but because MiniLM scored pembrolizumab vs Keytruda (same drug, brand vs generic) at 0.24 cosine similarity. Same drug, treated as unrelated. Pick five domain synonym pairs and check cosine. Most leaderboard wins die in production for this exact reason.

Principle 02

Defense in depth beats single-point precision.

My RAG system runs three independent guardrails: similarity threshold, LLM-as-judge document grader, and a generation-prompt refusal phrase. Each gate is allowed to be lenient because the composite is strict. Catastrophic out-of-domain queries fail at gate one; topically-adjacent-but-irrelevant queries (the 0.91 cosine hard negatives threshold alone misses) fail at gate two; insufficient context fails at gate three.

Principle 03

Library defaults aren't methodology.

My patient risk stratification project includes a causal inference layer on the NHEFS smoking-cessation cohort. The DoWhy library default produced +2.04 kg vs the Hernán-Robins reference of +3.4 kg — a 40% bias. Rebuilding per the methodology spec (quadratic confounder terms, logit propensity model, 1:1 nearest-neighbor matching with a 0.2 SD caliper, without replacement) put my estimate within 0.2 kg of reference. Three-way convergence with G-computation and IPW confirmed it.

Principle 04

The platform isn't the product, but the platform matters.

My pharma compliance project runs the same feature pipeline across Pandas, Snowflake, and PySpark with row-level equivalence verified. Real ML systems migrate across platforms as they scale. Proving the architecture survives that migration is part of the design, not an afterthought.

Projects

Three end-to-end systems, all live.

Project 01 · Clinical AI

Clinical Trial Intelligence System

Agentic RAG with a cross-encoder reranker for oncology and cardio-metabolic trial Q&A.

A natural-language interface over 484 ClinicalTrials.gov oncology and cardio-metabolic trials (3,264 chunks). PubMedBERT for domain embeddings, FAISS for similarity search, Groq Llama 3.3 70B for generation. Three architectural iterations: baseline RAG → agentic Corrective-RAG via LangGraph (LLM-as-judge document grading, query rewriting, bounded retries) → cross-encoder reranker (MS-MARCO MiniLM-L-6-v2) that halved in-corpus refusal rate from 67% to 33% on the 6 in-corpus queries within a 25-query golden set. Three independent guardrails for safety. Graceful rate-limit error handling on the live demo.

StackPython · LangGraph · Sentence Transformers · PubMedBERT · Cross-Encoder · FAISS · Groq Llama 3.3 70B · RAGAS · Streamlit

Live demo GitHub

Project 02 · Clinical ML & Causal

Patient Risk Stratification Pipeline

Production XGBoost + causal inference + survival analysis, deployed live on AWS Lambda.

End-to-end ML pipeline on the UCI Heart Disease dataset (303 patients, 13 clinical features). MLflow-tracked comparison across 5 model families surfaced XGBoost at 0.95 AUC, 0.93 recall at the clinical operating threshold, Brier 0.092 with documented calibration analysis. SHAP TreeExplainer for population and per-prediction explainability. Phase 2 extension to causal inference and survival analysis on the NHEFS cohort: propensity score matching + G-computation with DoWhy showed three-way convergence within 0.2 kg of the Hernán-Robins reference; Cox PH and Random Survival Forest at 0.80 concordance revealed pure age confounding behind a significant unadjusted log-rank. Deployed as a REST API on AWS Lambda via ECR + Mangum with FastAPI Swagger docs.

StackPython · Scikit-learn · XGBoost · PyTorch · SHAP · DoWhy · lifelines · scikit-survival · MLflow · FastAPI · Docker · AWS Lambda + ECR

Live API (Swagger) GitHub

Project 03 · Compliance Analytics

Pharma Compliance Spend Analytics

Anomaly detection on 16M+ real CMS Open Payments records, ported across three compute backends.

Real CMS Open Payments data — 16M+ records, $13B in payments. Sampled to 989K transactions across 289K healthcare providers for iteration speed. Five HCP-level features. Two-detector anomaly layer (within-specialty z-scores + global IQR with $500 monetary floor) plus an Isolation Forest ML layer on top. Rules-vs-ML reconciliation surfaced three compliance archetypes no single method catches alone — captured specialists, generalists, industry consultants. Same feature pipeline ported across Pandas (laptop), Snowflake (5 SQL queries reproducing EDA), and PySpark (DataFrame + Spark SQL) with row-level equivalence verified. Surfaced finding on real federal data: the top 4 device manufacturers — Arthrex, Stryker, Zimmer Biomet, Smith+Nephew — capture 59.7% of orthopedic surgery payments.

StackPython · Pandas · Scikit-learn · PySpark · Snowflake · Tableau

Tableau dashboards GitHub

Experience

Pharma and finance, end to end.

Sept 2022 — May 2026 · Pharma, Compliance Analytics

Amgen

Data Scientist

Architected the CMS Open Payments federal submission ETL (~509K records, ~$630M in the latest cycle) on AWS + Databricks across 4 annual submission cycles. Built multivariate risk scoring across 65+ ex-US countries and 40+ US products presented to Directors and Executive Directors; surfaced 3 high-risk markets undetected by the prior model. Developed statistical flagging across 5 transaction-monitoring processes. Led an NLP initiative that cut manual document categorization by 25%; prototyped LLM applications across 2 company-wide GenAI hackathons. Authored end-to-end documentation of three concurrent analytics workflows, adopted as team reference.

Jan 2022 — May 2022 · Internship

Nugget.ai

Data Scientist Intern

Built ML classification models that improved baseline matching accuracy by 15%. Engineered 20+ features across structural, linguistic, behavioral, and content-based categories.

Feb 2018 — Dec 2019 · Financial Services, Risk Modeling

Atos Syntel — at American Express

Data Scientist

Built risk scoring pipelines on 10M+ transactions covering spending velocity, geographic anomalies, behavioral segmentation. Implemented the team's A/B testing framework — sample sizing, randomization, power analysis, multiple-testing correction — across 12+ cycles per year for two years. Executive Tableau dashboards consolidating 5+ source systems that informed $2M+ in risk decisions.

Mar 2016 — Jan 2018

Atos Syntel — at American Express

Data Analyst

Production SQL workflows processing 5M+ transactions monthly with 99.5% accuracy. Automated weekly and monthly portfolio reporting, freeing 30+ hours/month of manual analyst time.

Education

May 2022

M.S. Data Analytics

Clark University

Feb 2016

B.E. Electronics and Telecommunication Engineering

Sant Gadge Baba Amravati University

Contact

Let's talk.

If your team is hiring — or if you want to compare notes on agentic RAG, causal inference, or shipping ML in regulated industries — reach out.

shrikant.sharma@myyahoo.com

linkedin.com/in/shrikant-sharma

GitHub

github.com/Shrikant-Sharma

Resume

Download PDF