Surya Kant Swain

At Wells Fargo, we had to choose among Gemini, GPT-4, and Bedrock models, but public leaderboards offered little insight into how any of them would handle our internal banking documents. I built a benchmarking pipeline that runs every model version against our own datasets and scores it on accuracy, schema compliance, and latency, and a CI check makes sure each endpoint is up before a full run starts. We now pick models based on our own data, and if a provider quietly updates a model, the problem shows up during a test run rather than in front of a user.

Every team at Wells Fargo that worked with term sheets, risk reports, or financial statements was writing its own parser, and plain text extraction kept losing the table layouts that give those numbers their meaning. I built a FastAPI service that accepts PDF, DOCX, PPTX, and scanned files from any source, then sends each file to Azure Document Intelligence for OCR and structure, or to PyMuPDF when speed matters more. Teams now integrate once instead of keeping their own parsers, and because the output keeps tables and sections intact, the search and retrieval tools built on top of it work better.

I started this personal project because most agent frameworks hide their logic inside chains of prompts, so when an answer is wrong, it's nearly impossible to tell which step caused it. I built the agent in LangGraph as an explicit graph, with a self-check step that rates how well the sources support an answer and goes back to retrieval when that support is weak. Every run saves a step-by-step trace, so when something breaks, I can find the exact step and replay it.