Technical Lead · Perfios Software Solutions · Dehradun, India
Hi, I'm Pankaj Chauhan
I lead teams that turn millions of messy PDFs into clean, structured data - and trade screens for ridgelines on the weekend.
// about
Two worlds,
one approach.
"Be the change that you want to see in the world."
I build scalable backend systems for fintech, where my work revolves around one messy problem: turning millions of complex PDFs and images into clean, structured data. I've built the core pipelines that extract at massive scale, engineered custom parser frameworks, and shipped fraud-investigation workflows that have reviewed over 25 million documents.
Lately my focus has shifted toward applied AI inside enterprise systems - LLMs, RAG, and agentic workflows powering risk-monitoring products that are smarter and far cheaper to run. Combining traditional data mining with modern generative models is exactly the kind of problem I want to keep solving.
Away from the keyboard I'm usually at altitude. Long approaches and quiet summits in the Himalaya are how I reset - and planning an expedition turns out to be a lot like planning a release.
// experience
Where I've shipped.
$ git log --oneline --graph --decorate
- + Leading parser frameworks, prompt engineering, and production ML best practices while driving AI workflow adoption across products — reducing manual work and accelerating feature delivery.
- + Redesigned a legacy document-OCR service into a modular, multi-engine pipeline (Google Vision / Tesseract, Python) — introduced streaming batch processing that keeps memory constant regardless of document size, parallelized OCR execution, and eliminated a redundant preprocessing stage consuming ~95% of pipeline CPU. Delivered 70% faster processing, 34% lower peak memory, and 41% less CPU on production.
- + Architected a Corporate Risk platform monitoring Indian entities for adverse media and governance risk — a 10K/day pipeline with custom ML noise-filters and a scalable API layer.
- + Cut LLM costs by 85% through advanced prompt engineering while accurately classifying Fraud, Litigation, and Misconduct risk.
- + Built a DGFT trade-risk pipeline that analyses daily court judgments, combining multi-stage OCR with prompt engineering to classify complex legal risk.
- + Led research on table extraction from PDFs and images, shipping a versatile solution that has supported extraction across 100M+ documents.
- + Developed an industry-first PDF Tampering Detection solution for banking fraud, analysing 1.5M+ documents with an automated scoring mechanism that slashed false positives.
- + Built a versatile parser-utility framework for fields, tables, and key-value pairs from PDFs, images, and OCR - cutting development time 80-85% at 98.6% accuracy.
- + Improved accuracy and latency with OCR error management, caching, and region-of-interest cropping.
- + Built low-latency parser APIs for KYC services, integrating Google OCR and Azure OCR.
- + Reduced API latency 40% via caching and optimised search algorithms.
- + Created custom parsers for complex tables, removing the need for expensive third-party APIs.
- + Wrote SQL and MongoDB migration scripts and built front-end features with React and jQuery.
- + Took ownership from root-cause analysis to production release.
// toolkit
What's in the stack.
Languages 4
Generative AI & RAG 9
Data Mining 8
Web & Data 6
Tooling 6
// certifications
Staying sharp.
// education
Where it started.
National Institute of Technology, Srinagar
B.Tech · Information Technology
J&K, India