Harish Janagam

Project Summary: Full-stack Agentic AI platform targeting legal, procurement, and delivery teams handling IT contracts (SOW, SOWA,
SLA, MSA). Designed for production-grade reliability with enterprise-level observability and zero-trust document integrity.
Tech Stack: React + TypeScript + Vite
• FastAPI
◦ CrewAI
◦ Docling
◦ PostgreSQL (asyncpg + pgvector)
◦ Qdrant
◦ OpenAI GPT-4o
◦ Celery
◦ Alembic
• Docker
▸ Architected a CrewAI-based hybrid multi-agent orchestration layer with specialist agents for clause extraction, risk scoring,
obligation mapping, and anomaly detection
▸ Designed a 4-agent crew (Contract Analyst, Risk Assessor, Obligation Mapper, Report Generator) with inter-agent task delegation
and structured handoff protocols
▸ Built a dual-store retrieval architecture linking PostgreSQL clause UUIDs with Qdrant HNSW vector indexes — enabling sub
second semantic search with no extra DB roundtrips
▸ Implemented a custom ContractID schema (YYYYMMDD_HHMMSS-) with 100% verbatim clause storage for
audit integrity and tamper detection
▸ Parallelised OpenAI extraction across chunked clause batches to process 500-page contracts in under 60 seconds end-to-end
▸ Replaced unstructured.io with Docling for local document parsing (zero cloud cost, zero data egress) supporting PDF, DOCX, and
multi-column layouts
▸ Resolved enterprise TLS inspection failures by injecting OS certificate store via the truststore library; suppressed CrewAI OTEL
telemetry noise in corporate environments
▸ Designed Azure cloud deployment architecture (Azure Container Apps, Azure Document Intelligence, Azure Cosmos DB) with
detailed cost modelling across three deployment tiers
▸ Built React + TypeScript + Vite frontend with contract upload, multi-tab clause explorer, risk dashboard, and obligation timeline
views Alight (Healthcare Analytics)

Project Summary: End-to-end Agentic Contract Analysis Platform for a leading US healthcare PBM. Enables legal teams to upload,
compare, and validate contract documents with AI-driven term extraction and structured feedback workflows.
Tech Stack: Python
• FastAPI
◦ LangGraph
◦ LangChain
◦ Azure OpenAI (GPT-4o/4o-mini)
◦ Azure Document Intelligence
◦ Azure AI Search
◦
MongoDB
• GraphQL
◦ Pydantic
▸ Designed and deployed a multi-agent LangGraph pipeline with dedicated agents for contract term extraction, email notification,
and feedback control
▸ Built specialised tool functions integrated with agents for clause-level extraction across PBM pricing tables and contractual
definitions
▸ Orchestrated stateful LangGraph nodes with conditional edges for dynamic routing between extraction, validation, and user
notification phases
▸ Achieved 100% accuracy on term and pricing table extraction by combining Azure Document Intelligence layout analysis with
GPT-4o structured outputs
▸ Implemented an email notification sub-agent within the graph to deliver real-time alerts to content writers and auditors upon
document processing
▸ Designed human-in-the-loop feedback node allowing legal reviewers to flag, correct, and re-inject validated terms back into the
pipeline
▸ Applied Azure AI Search for vector embeddings and similarity retrieval across large contract corpora; MongoDB for persistent
clause and audit storage

Project Summary: Predictive modelling on patient diagnosis data for 50+ major US healthcare clients including UHG, BCBS, Cigna, and
Athena. Built hyper-personalisation and classification models on insurance, claims, and 401K datasets at scale using Databricks and
Spark.
Tech Stack: Python
• Databricks
◦ Apache Spark
◦ PySpark
◦ SparkSQL
◦ MLlib
◦ PyTorch
◦ TensorFlow
◦ AWS SageMaker
◦ S3
◦ Hue/Impala
◦
LangChain
• HuggingFace Falcon
▸ Engineered large-scale data pipelines on Databricks using PySpark and SparkSQL to process and merge multi-source healthcare
datasets (claims, diagnosis, 401K) across distributed Hue/Impala databases and S3
▸ Performed complex Spark transformations — joins, window functions, aggregations, and partitioning — on 50+ client datasets to
produce hyper-personalised feature matrices for downstream ML models
▸ Optimised Spark job performance through broadcast joins, caching strategies, and partition tuning, reducing pipeline runtimes
significantly on large patient cohort data
▸ Built and deployed PyTorch and TensorFlow deep learning models for patient classification and risk stratification on structured
clinical and insurance data
▸ Developed and deployed multiple classification and regression models (XGBoost, LightGBM, Random Forest) on patient
diagnosis, claims, and new-hire insurance datasets
▸ Built a GenAI Q&A pipeline using LangChain over PDF clinical data; created a HuggingFace Falcon-powered text generation model
for internal knowledge retrieval
▸ Implemented NER-NLP pipelines for healthcare text classification, extracting structured entities from unstructured clinical notes
▸ Managed end-to-end ML lifecycle on AWS SageMaker: data wrangling, feature engineering, model training, deployment, and
monitoring across 50+ healthcare client environments

▸ Created supply-and-demand ML models for BmaaS, MonaaS, DBaaS, and CaaS service pricing prediction
▸ Dockerised microservices and established pipeline architecture following MVC patern for multi-service orchestration
▸ Migrated aiohtp-based services to Flask framework; built RESTful APIs to ingest real-time service telemetry (Financial

Stack: Python
• Pandas
◦ ETL
◦ SSMS
◦ Jupyter
▸ Executed content migration from legacy to updated financial data services with automated data comparison and validation
pipelines
▸ Extracted and transformed structured financial datasets for production deployment across FactSet content systems

Python Pytesseract
• Scikit-learn
◦ Linear Regression
◦ Pandas
◦ Jupyter
▸ Developed regression models to predict soap bar hardness across chemical compositions and temperature conditions, providing
R&D insights for multiple product brands
▸ Analysed how different chemical parameters react under varying weather conditions to model soap bar behaviour at scale
▸ Automated data extraction and consolidation from multi-sheet Excel research files, cleaning and aligning data to business
requirements
▸ Built an invoice image parsing pipeline using Pytesseract to extract structured fields (quantities, values, dates) from scanned
invoice images
▸ Applied Linear Regression and statistical methods for outlier detection, missing value imputation, and predictive quality scoring
on R&D datasets

• Scikit-learn
◦ Power
◦ SQL
▸ Analysed hotel and resort booking data to uncover customer behaviour paterns and provide business intelligence to account
managers
▸ Developed regression models for customer booking rate prediction; visualised insights in Power BI dashboards for executive
presentations
▸ Collaborated cross-functionally with data science and business stakeholders within a banking application (Finacle) context