Building
AI + data products
Agentic workflows, decision support, and interfaces that make complex systems feel calm.
MS Information Management @ UIUC · Data Science @ COUNTRY Financial
I build intelligent products with data + AI.
I'm Pooja — a data scientist with a product mindset, obsessed with making AI genuinely useful, not just impressive.
AI Engineer, Data Scientist, Forward Deployed Engineer, Solutions Analyst, Machine Learning Engineer, Technical Product ManagerGraduating May 2027
About / currently
A concise picture of how I work and what I am building toward.
Pooja Sahu combines AI curiosity, data discipline, and product thinking to move from interesting ideas to trustworthy systems.
Learning
Production AI systems
Evaluation, retrieval, deployment patterns, and the tradeoffs behind robust ML delivery.
Exploring
AI Engineer / Applied AI / Data Scientist roles for 2027
Teams where product judgment and technical depth both matter.
Experience
Where I've built, shipped, and learned.
2026Current role
NowCOUNTRY Financial
Data Science Intern/May 2026 – Present/Illinois, USA
Current role
NowCOUNTRY Financial
Data Science Intern/May 2026 – Present/Illinois, USA
Building LLM evaluation, document intelligence, and large-scale data pipelines for enterprise insurance AI.
- Developed the PolicyPro multi-layer LLM evaluation framework, combining deterministic validation, semantic similarity, evidence-based assessment, and risk classification to reduce manual review of AI-assistant outputs
- Built an AI-powered document validation system using OCR, PDF layout analysis, and template comparison to automate insurance billing QA checks
- Developed memory-efficient preprocessing and feature-engineering pipelines for 14.5M+ insurance records supporting actuarial modeling
2023Enterprise technology
Accenture
Security Delivery Analyst/Jun 2023 – Jul 2025/Mumbai, India
Enterprise technology
Accenture
Security Delivery Analyst/Jun 2023 – Jul 2025/Mumbai, India
Automated enterprise security and risk workflows and built the data validation behind them.
- Configured ServiceNow IRM/SecOps workflows that cut manual review effort by 20% (~200 hours/month)
- Built Python data validation and anomaly-detection workflows across 50K+ enterprise records, improving accuracy and SLA adherence
- Engineered integrations, scheduled data feeds, and testing across ServiceNow and RSA Archer
- Collaborated across 3 engineering squads on requirements, testing, and release planning in an Agile environment
2021Computer vision
Capgemini
Data Science Intern/May 2021 – Jun 2021/Mumbai, India
Computer vision
Capgemini
Data Science Intern/May 2021 – Jun 2021/Mumbai, India
Built a real-time weapon detection model and automated reporting validation.
- Developed a YOLOv5 weapon detection model trained on 3,000+ custom-labeled images, reaching 98.48% classification accuracy
- Automated data-validation and reconciliation workflows in Python, reducing recurring reporting errors by 15–20%
Selected work
Systems I've designed across AI, ML, and data.
Open any project for the approach, the decisions behind it, and the stack.
01LLM evaluation / enterprise AIPolicyPro LLM Evaluation Framework
A multi-layer evaluation framework that systematically flags response-quality issues in an enterprise AI assistant and cuts manual review effort.
01LLM evaluation / enterprise AI
PolicyPro LLM Evaluation Framework
A multi-layer evaluation framework that systematically flags response-quality issues in an enterprise AI assistant and cuts manual review effort.
- Layers deterministic validation, semantic similarity, and evidence-based assessment
- Risk classification to prioritize the responses that need human attention
- Automated testing of LLM response quality and reliability at scale
LLM EvaluationSemantic SimilarityRisk ClassificationPython
02Multi-agent systems / generative AIsust-AI-naible
A multi-agent, carbon-aware compute scheduler where specialized agents collaborate to recommend when and where workloads should run.
02Multi-agent systems / generative AI
sust-AI-naible
A multi-agent, carbon-aware compute scheduler where specialized agents collaborate to recommend when and where workloads should run.
- Grid Forecast, Scheduler, Workload, and Policy agents with distinct responsibilities
- Orchestration logic for agents to exchange context and reach a joint recommendation
- Scheduling decisions shaped by carbon intensity, not just capacity
PythonMulti-Agent SystemsGenerative AIOrchestration
03RAG / conversational AISRB RAG-Powered WhatsApp Assistant
A retrieval-augmented chatbot that answers on WhatsApp with grounded responses drawn from an indexed knowledge base.
03RAG / conversational AI
SRB RAG-Powered WhatsApp Assistant
A retrieval-augmented chatbot that answers on WhatsApp with grounded responses drawn from an indexed knowledge base.
- FastAPI backend with FAISS vector search over embedded knowledge
- WhatsApp messaging through Twilio for a zero-install user experience
- Retrieval and conversational workflow connecting user queries to relevant context
PythonFastAPIFAISSTwilioEmbeddings
04Machine learning / data pipelinesLarge-Scale Insurance Data Processing & Risk Modeling
Memory-efficient pipelines that turn 14.5M+ insurance records and 180+ variables into model-ready data on local compute.
04Machine learning / data pipelines
Large-Scale Insurance Data Processing & Risk Modeling
Memory-efficient pipelines that turn 14.5M+ insurance records and 180+ variables into model-ready data on local compute.
- Chunk-based preprocessing and feature engineering for datasets that don't fit in memory
- Automated data-quality checks, special-value handling, and variable selection
- A reusable pipeline for downstream actuarial modeling and analytics
PythonMachine LearningData Pipelines14.5M+ Records
05Computer vision / deep learningReal-Time Weapon Detection
A YOLOv5 object-detection model for security surveillance, fine-tuned on custom-labeled images to 98.48% classification accuracy.
05Computer vision / deep learning
Real-Time Weapon Detection
A YOLOv5 object-detection model for security surveillance, fine-tuned on custom-labeled images to 98.48% classification accuracy.
- Trained and fine-tuned on 3,000+ custom-labeled images
- Real-time detection pipeline built for surveillance use cases
- 98.48% classification accuracy
YOLOv5Computer VisionPythonObject Detection
Capabilities
What I bring to an AI team.
LLM & agentic systems
Built PolicyPro, a 4-layer LLM evaluation framework for an enterprise AI assistant, plus a multi-agent scheduler and a RAG assistant on WhatsApp.
LangChain · LangGraph · LlamaIndex · MCP · FAISS · NeMo Guardrails
Machine learning
Fine-tuned a YOLOv5 detection model on 3,000+ labeled images, reaching 98.48% accuracy for real-time surveillance.
scikit-learn · XGBoost · YOLOv5 · Feature engineering · Cross-validation
Data engineering
Built memory-efficient pipelines that make 14.5M+ insurance records and 180+ variables model-ready on local compute.
Python · SQL · Pandas · PySpark · Cloudera · GCP
Enterprise delivery
Automated ServiceNow risk workflows at Accenture, saving ~200 hours a month of manual review across 3 engineering squads.
ServiceNow · RSA Archer · Git · Power BI · Tableau
Education
Engineering roots, information-science focus.
UIUC
GPA 4.00 / 4.00Aug 2025 – May 2027
University of Illinois Urbana-Champaign
School of Information Sciences
MS in Information Management
Relevant coursework9
Fall 2026In progress
IS 514
Applied Business Research
CS 409
The Art of Web Programming
INFO 490
Advanced AI Web Applications
Spring 2026
IS 455
Database Design and Prototyping
IS 492
Introduction to Generative AI
IS 534
Information Consulting
Fall 2025
IS 504
Sociotechnical Information Systems
IS 507
Data, Statistical Models and Information
IS 525
Data Warehousing and Business Intelligence
NMIMS
GPA 3.53 / 4.00Jul 2018 – May 2023
NMIMS
Narsee Monjee Institute of Management Studies · Integrated dual degree
MBA in Technology Management
Major in Business Intelligence & Analytics · Minor in Marketing Analytics
B.Tech in Electronics & Telecommunication
Beyond work
Strength, new places, and a little adventure.
The habits that keep me grounded off-screen shape how I show up on every team.
Discipline
Gym, Pilates & Climbing
Lifting, reformer Pilates, and the climbing wall — where I practice consistency: small, measurable progress, repeated until it compounds.
Perspective
Travel
Lakeshores, forest trails, and new horizons — travel keeps me curious and reminds me to look at the bigger picture.
Contact
Have an interesting problem to solve? Let's talk.
Recruiters, hiring managers, and collaborators can reach me directly without friction.