Dhiraj(Raj)Jha

Five years of sensitive Tier 3 data was trapped in a legacy system, scattered across PDFs, scanned files, and Word documents, and it needed to migrate to a modern third-party platform (Onsite Systems) without loss or corruption. The data had no consistent structure, and most of it couldn't move without being digitized first. I built an ETL pipeline using a multi-engine parser (PyMuPDF and pdfplumber running in parallel) with automated divergence flagging for when their outputs disagreed. For unstructured anomalies the parsers couldn't confidently resolve, I integrated a locally-hosted Llama 3 instance via Ollama, which kept sensitive data off external networks. The pipeline normalizes everything into JSON/YAML payloads structured for the vendor's strict API schema. Manual data entry was cut by 90%.

Google's PRICE (Production Risks in Core Engineering) program periodically assesses system health across production infrastructure, but every cycle was entirely manual. Each assessment took around three days and required engineers to handle emailing, reporting, and bug tracking by hand. I was tasked with automating it end to end. I built PRICELess, a tool that orchestrates Google Workspace APIs to handle all communication and reporting automatically, and layered in a RAG pipeline that grounds LLM outputs in SRE principles so assessment reports come out standardized and citation-backed rather than freeform. I also built an auto-authentication service using LOAS to resolve the cross-service permission failures breaking calls between microservices. In the end, assessment time dropped from three days to fifteen minutes, and every previously manual step was eliminated.

No existing travel tool generates a complete, optimized itinerary deterministically. They surface options, not answers. Offtofly is built to solve that. I am leading a five-person engineering team building the system from scratch as distributed microservices in Go, Python, and C++, communicating over REST and gRPC. The core is a custom graph database engine that models experiences as nodes with typed edges like location, timing, pace, and budget, so the constraint solver can compose valid itineraries without delegating the planning logic to an LLM.