Richy Koshy

MAX AI, enterprise RAG platform. Owned the full Python backend across three Azure services; now PRMI’s primary knowledge-retrieval tool organization-wide.

• Retrieval: hybrid vector + semantic ranking on Azure AI Search with HyDE as a fallback when recall is low; Azure OpenAI GPT-4o with ada-002 embeddings.

  One index per project, with multi-tenant access mapping in SQL Server behind Entra ID auth.

• Ingestion: Durable Functions pipeline, idempotent via progress-log checkpointing with per-file failure isolation.

  Local PDF partitioning and table extraction with unstructured (poppler + Tesseract), shipped as a Docker image on Azure Container Registry.

• Chat: FastAPI service streaming NDJSON with metadata-first framing, so citations render before the first token.
• Utility APIs: a Function App for auth, projects, users, session history and feedback.

Read the MAX AI deep dive

Final Docs refactor. The automation governs how PRMI assembles and delivers post-closing packages to its investor banks, where delays put funding lines and regulatory standing at risk.

• Approach: rebuilt a monolithic automation into config-driven dynamic dispatch: one YAML config per investor, and a dispatch table for the few investors that need custom handlers, treating each investor as data, not code.
• Result: 4 new investors onboarded in 1 month vs. a full quarter before; onboarding cut from 2-4 weeks to days; adopted as the template for 2 other initiatives.

Read the Final Docs deep dive

Encompass REST API SDK. Internal Python package consumed by all 13 automations:

• OAuth2 token caching
• concurrent paginated pipeline reads (ThreadPoolExecutor)
• transactional loan writes with a lock → write → read-back → unlock pattern that prevents lost updates on shared loans
• eFolder document CRUD, disclosure tracking, and 429 retry with backoff.

Read the Encompass deep dive