Bijaya Kumar Pariyar

What started as curiosity about data turned into a full journey through the ML and AI ecosystem. I went from training my first Random Forest to designing production-grade RAG systems and multi-LLM orchestration layers, without a formal CS degree, driven entirely by self-study and shipping real projects. When I believed I was ready, I reached out to a US-based commercial real estate fintech directly. They gave me a shot, and I made it count.

During the internship I built the core of the core AI platform for commercial real estate lending, authoring 150+ commits to the core service, roughly 70% of its recent production-branch history. The platform runs a RAG pipeline that classifies and extracts structured data from 20+ financial document types in 55–90 seconds end to end, multi-provider LLM orchestration (AWS Bedrock, Gemini, OpenAI) with fallback chains and a throttle-aware circuit breaker, and a deterministic credit-decision engine deliberately separated from LLM narrative generation, so model hallucination or prompt injection cannot alter a lending decision. Underneath sits semantic retrieval over pgvector with HNSW search and cross-encoder reranking, and a two-service async architecture deployed to AWS via CI/CD.

In parallel, I shipped six additional production services, including a solo-built payments integration with signed webhooks, an internal knowledge-base RAG service, and an intake-to-offer valuation engine, and contributed to SOC 2 and NIST 800-53 readiness, designing the control matrix and audit-trail documentation. Independently, I also built a semantic code-search tool over a 42-repository organization, using tree-sitter AST chunking, local embeddings, Postgres/pgvector, and a custom MCP server integration.

Underneath the production work is a deep classical-ML foundation: 20+ structured end-to-end projects across classification, regression, time series forecasting, clustering, and anomaly detection, including 1M+ rows for fraud detection and 1.7M+ rows for price forecasting. My approach combines rigorous evaluation, model explainability through SHAP, and reproducible research, including a published research preprint on SHAP-based feature selection.