Pavan – LLM, RAG, LangChain

An intake automation platform for senior care facilities processing inbound medical referral packets from hospitals. Incoming PDFs are converted to text via document intelligence, then passed to an LLM with a strict output schema to extract patient demographics, diagnoses, current medications, biodata and insurance coverage. The facility uses the structured output to decide whether it can clinically support the patient and whether the placement is financially viable. The platform also surfaces facility capacity overview, insurance verification status and background checks. Built under PHI/PII constraints with role-scoped access enforced end to end.

• Designed the agentic layer of the platform, from schema definition through to deployment topology.
• Modelled the extraction contract as Pydantic schemas, enforcing structured LLM output over free-text parsing.
• Built the LangGraph agent and exposed it through FastAPI, serving it on AKS.
• Configured Helm charts and Spinnaker pipelines for deployment into the AKS cluster.
• Implemented access control with Microsoft Entra ID, scoping data visibility per facility role.
• Authored and tuned the extraction prompts against the required output schema.

A retrieval-augmented question answering system over an enterprise financial institution's insurance policy library, built so that in-house insurance advisors could answer client questions from source documents instead of manual lookup. Policy PDFs are converted to text, split with semantic chunking, embedded and indexed in a vector store. User queries are embedded and matched against the index, returning answers with citations back to the originating policy clause.

• Implemented semantic chunking and attached search metadata to each chunk to support filtered retrieval.
• Built and indexed the vector store backing the policy corpus.
• Designed and implemented the Python backend for the RAG search application, including the citation layer.
• Built an event-driven ingestion pipeline that automatically chunks and vectorises any new document landing in blob storage, removing manual re-indexing.

A centrally governed data and analytics platform for a major financial institution, giving internal analytics and data science teams self-serve access to lakehouse data under fine-grained, auditable permissions. The platform covers ingestion, cataloging, warehousing, and ML workspaces, with governance enforced as a service rather than per-team convention.

• Built governance patterns around Lake Formation to control access to Glue, Redshift, S3 and SageMaker from one permission model.
• Developed reusable CDK constructs for Lake Formation permissions, LF-tags and data sources.
• Implemented tag-based access control to enable fine-grained, attribute-driven permissions.
• Defined platform personas against which Lake Formation permissions were enforced.
• Designed a custom EventBridge-based logging solution for AppFlow, cutting debugging time from days to hours.
• Closed gaps in native Glue logging and reshaped the output into a consumable format.
• Led a cloud infrastructure optimisation initiative, producing RACI documents for all platform services and reducing team onboarding from weeks to days.
• Led the build of an operational monitoring and alerting framework for platform services.
• Migrated CDK deployments from Bitbucket to GitHub Actions and debugged every platform construct for a clean cutover.
• Led a team of 5 on the data/AI platform initiative for 6 months.

An applied data science engagement for a national revenue agency, targeting anomaly detection and case segmentation across taxpayer case files. Work covered a state-of-the-art survey of semi-supervised learning approaches for a labeling-constrained problem, followed by clustering and anomaly detection on production case data.

• Ran a comprehensive survey of semi-supervised learning methods against a labelling-constrained business problem.
• Executed segmentation using several advanced clustering algorithms.
• Implemented anomaly detection with isolation forest and LOF to flag potential data security threats.
• Trained predictive models for anomaly classification across case datasets.
• Automated case-file report generation in R, sharply reducing manual reporting time.
• Built a Git setup template for the agency intranet to establish version control practice across the team.