Production-grade AI systems, engineered end to end.
Architecting secure, deterministic AI pipelines for commercial fintech: bridging the gap between RAG experimentation and enterprise-scale reliability.
Where to go next
About
My background, career journey, research, and the values behind the work.
Projects
The production platform I built, plus 14+ open-source projects across RAG, ML, and NLP.
Skills
The languages, frameworks, and cloud infrastructure I build with, grouped by category.
Notes on RAG systems, ML experiments, and what I learn while building.
Hi, I'm Bijaya
I build production AI systems: retrieval pipelines, LLM orchestration, and the engineering that makes them dependable enough to sit behind a regulated decision.
On a Data Scientist internship at a US commercial real estate fintech, I built the core of an AI underwriting platform: RAG document intelligence over 20+ financial document types, multi-provider LLM orchestration with fallback chains, and a deterministic credit engine that keeps the model out of the decision itself. I also maintain edaprep, an open-source preprocessing library on PyPI with contributors of its own, on a foundation of 20+ end-to-end machine learning projects.
From notebooks to production
The kind of systems I design, ship, and own end to end.
Multi-LLM orchestration
Provider fallback chains across AWS Bedrock, Gemini, and OpenAI, with circuit breakers and graceful degradation.
RAG & retrieval
pgvector HNSW search, Titan and sentence-transformer embeddings, and cross-encoder reranking over real documents.
Deterministic decision engines
Rule-based credit risk scoring with zero LLM in the critical path: auditable, explainable, and reproducible.
Async cloud services
FastAPI and async job workers on AWS ECS Fargate, shipped via GitHub Actions CI/CD with layered security.