Saibal Roy

Engineering Manager · Building scalable solutions & teams

I've been building software for 13+ years, most of them at Techno Exponent, where I started as a founding engineer and grew into the Engineering Manager who runs cloud delivery. Today my DevOps team looks after engagements in legal technology, insurance technology and scientific research.

All of them run on one AWS deployment pattern: ECS and ECR, infrastructure as code, the same CI pipeline shape. So when a new engagement starts, we aren't designing from scratch. We're setting up something that already works in production.

I own the architecture, the delivery and the team end to end, and I'm the client's single point of contact. I'm still hands-on.

How I build

Outside client work, I build open source the same way: I start from a real production requirement and turn it into a reusable solution that's cheap to run and reliable enough to trust. The systems thinking is part of the deliverable:

Cost first: start from the unit economics and size down to the smallest machine that measurably works.
Reliability you can prove: an acceptance check for every requirement, and a go-ahead gate before anything ships.
Evidence before decisions: benchmarks, written down with the trade-off and who decided.
Secure and maintainable by default: localhost-only services, no stored credentials, pinned versions, one supported platform.

The prompts I build with · How docling-batch-extract was built · How I publish open source on GitHub

Projects

Each one starts from a production requirement and ships with its design history, benchmarks and test results.

docling-batch-extract

Batch PDF → Markdown for RAG search on one CPU-only server, using docling-serve. Built for legal and claims-insurance documents (including scans) as a low-cost alternative to cloud OCR APIs. Documents never leave your server.

CPU only · 2 vCPU / 8 GB target
Tested on Ubuntu LTS
MIT license

Documentation Source on GitHub