Sagar Virmani - Senior Data Scientist

🌟 About Me

Senior Data Scientist with 5+ years of experience building production ML and GenAI systems that deliver measurable revenue impact. My work spans the full lifecycle: from problem framing and model development through to MLOps infrastructure, LLM application engineering, and business deployment.

🤖 GenAI & LLM Engineering

LLM-powered conversational BI on GCP Vertex AI with MCP server layer - 85–90% accuracy validated
Gemini 2.5 Pro with GCS-backed JSON data models
Natural language to SQL across multi-database environments

⚙️ Production MLOps

End-to-end ML pipelines from research to production under SLA
Docker, Airflow DAGs, MLflow, GCP Vertex AI, AWS EC2, CI/CD
Drift monitoring, automated alerting, governance documentation, rollback capability

📈 Business Impact Focus

80%+ forecast error reduction in live sales systems
5.3% direct revenue increase via predictive forecasting
Production systems scoring 100,000+ records continuously

📈 Business Impact & Metrics

>80%

Forecast Error Reduction

+5.3%

Direct Sales Increase

100K+

Leads Scored in Production

90%

Clinical Detection Accuracy

3

BigQuery DBs on Single Chat Interface

5+

End-to-End Production Deployments

🚀 Featured Projects

Built & Accuracy-Validated · 85–90% Accuracy

🤖 Conversational LLM-Powered Business Intelligence System

GenAI / LLM Application · Natural Language → Instant Data Insights · Gemini 2.5 Flash & Pro on GCP Vertex AI · MCP Server Layer

How it works

User query enters via enterprise frontend → FastAPI orchestrates all ingress and egress → Intent classifier routes to deterministic SQL templates or Gemini 2.5 Flash for dynamic generation → MCP layer enforces responsible AI controls and loads GCS data models → BigQuery executes SQL across 3 enterprise databases → Gemini Answer converts the output table into a natural language response returned to the frontend.

Gemini 2.5 Pro / Flash GCP Vertex AI BigQuery FastAPI MCP Server Layer GCS · JSON Data Models Python

Innovation: MCP server layer intercepts all LLM tool calls before BigQuery execution, validating SQL field names and blocking hallucinated identifiers. GCS-backed JSON data models encode cross-database schema relationships and business semantics across 3 BigQuery datasets.

Accuracy: 85–90% achieved across Gemini 2.5 Flash and Pro on a representative business query set. Flash validated for production-scale deployment at lower cost without meaningful accuracy loss.

Impact: Zero SQL knowledge required. Business users query all 3 databases from a single chat interface. System built and validated on GCP Vertex AI

Explore Case Study →

Independent Project · 100% Tool Call Accuracy

🤖 Patient Follow-Up Agent — Agentic AI

Agentic AI · LangChain 1.x · LangGraph · Claude Haiku · ReAct Pattern · FastAPI

How it works

Clinical query enters the system → LangChain agent (create_agent, LangGraph orchestration) reasons about which tool to call using the ReAct pattern → agent autonomously calls 4 custom tools: load patient record, assess clinical risk (BNP, HbA1c, eGFR, BP, PHQ-9 thresholds), generate care plan, detect missed appointments → Claude Haiku reasons over tool outputs and decides next steps → structured clinical response returned via FastAPI REST endpoints.

LangChain 1.x LangGraph Claude Haiku ReAct Pattern FastAPI Python Healthcare AI

Architecture: Single-agent, tool-calling system using LangChain 1.x create_agent with LangGraph state graph orchestration under the hood. No RAG, no vector databases — purely agentic tool orchestration with ReAct reasoning loop.

Results: 100% tool call sequence accuracy and 100% task completion across all test scenarios on 100-patient synthetic dataset. Identified 15 missed appointments and 8 high-risk heart failure patients (BNP > 900 pg/mL).

Evaluation: Process-level metrics and output grounding spot-checks — a deliberate design choice given no clinician-labeled ground truth was available. Built independently to deepen hands-on agentic AI expertise.

Explore Case Study →

Production · Live on GCP Vertex AI

🎯 MLOps Sales Pipeline Intelligence System

Production MLOps · Sequence Modelling · >80% Forecast Error Reduction

Hidden Markov Model GCP Vertex AI Apache Airflow Docker Salesforce BigQuery Python

How it works

Salesforce stage progression data is extracted and filtered for directional transitions → HMM models each deal as a sequence of hidden states to generate closure probability signals → pipeline containerised in Docker with train, evaluate, and score CLI commands → Airflow DAGs orchestrate scheduling and pipeline execution on GCP Vertex AI → scores written to BigQuery with full governance documentation, audit trail, and rollback capability.

Innovation: Hidden Markov Models capture temporal stage transition sequences to generate 12-month advance closure probability signals that static models miss entirely

Impact: 80% forecast error reduction, 5x more accurate than experienced sales reps, continuously scoring 100,000+ leads in production on GCP Vertex AI

Explore Case Study →

Production · SLA-Based Quarterly Delivery · GCP

📊 AI-Powered Sales Forecasting Engine

Time-Series Forecasting · +5.3% Sales Increase · 60% Forecast Error Reduction

ARIMA & SARIMAX PyCaret GCP Python · Pandas Statsmodels

How it works

Top 10 suppliers analysed across 10 years of pharmaceutical sales data → ARIMA/SARIMAX models trained per supplier, with order parameter tuning to correct historical anomalies and capture seasonality patterns → automated back-testing and evaluation across model configurations via PyCaret → 12-month forecasts generated for top 3 suppliers on GCP under quarterly SLA → outputs delivered with supplier margin intelligence for procurement and commercial decisions.

Innovation: ARIMA/SARIMAX models trained per supplier on 10 years of pharmaceutical data; ARIMA order tuning corrected historical anomalies distorting forecasts, SARIMAX captured supplier-specific seasonality patterns

Impact: 5.3% direct sales increase, MAE reduced from 10-15% to 4-6%, plus actionable supplier margin intelligence for procurement decisions

Explore Case Study →

Deployed · Live on AWS EC2

🏥 AI-Powered Thyroid Clinical Decision System

Healthcare AI · Explainable ML · 90% Detection Accuracy · 0.9 ROC-AUC

XGBoost Random Forest KNN SMOTE scikit-learn Flask Docker AWS EC2

Innovation: Explainable classifier surfaces top contributing clinical markers alongside every prediction, giving clinicians the transparency to validate the AI recommendation before acting on it

Impact: 90% detection accuracy, 0.9 ROC-AUC, 40% improvement over previous diagnostic methods, 10% faster diagnosis

Explore Case Study →

🔧️ Technical Stack

🤖 GenAI & LLM Engineering

Gemini 2.5 Pro / Flash GCP Vertex AI LLM Prompt Engineering Text-to-SQL JSON Data Models GCS Bucket FastAPI RAG Pipelines LangChain HuggingFace Transformers Context Window Management