About

Hi — I’m Sounak, a backend and infrastructure engineer based in Bangalore. I work at the intersection of applied AI and backend systems — building the kind of plumbing that turns probabilistic models into systems a business can actually run on.

Right now I’m at Staple — an AI document-processing platform where we extract structured data from invoices, contracts, and claims with cryptographic audit trails, so regulated teams can defend every figure to an auditor. Day-to-day that means owning the message-based ingestion that moves documents through the stack, the LLM-driven detection that decides what each one is, and the redaction systems that handle hundreds of thousands of client documents.

Before Staple I was at Pratilipi, a self-publishing platform for stories in Indian languages, where I spent two years as a data scientist. I owned recommendations (50% lift in reads, 4× subscriptions), shaved 70% off the ML deployment cost, and stood up most of the internal ML platform on AWS. Pratilipi taught me what ML feels like in production — latency, cost, observability, the long tail of inputs nobody trained for.

The thread between the two is the same one I keep getting nerd-sniped by: AI in production. Demos are easy; the interesting work is in evals, retries, schemas, audit trails, and the boring connective tissue that decides whether a probabilistic system can be trusted with somebody’s document — or somebody’s money. I’m building the part of AI that doesn’t get demoed: the agents, tools, memory layers, and pipelines that turn an impressive model into a system that quietly runs a business.

When I’m not at a keyboard I’m probably out with a camera, lifting, or building small tools nobody asked for.

Work

Staple · Software Engineer · 2024–present

Building infrastructure for a cognitive document-processing platform.

  • Engineered a RabbitMQ-based ingestion system — the backbone for high-volume document processing.
  • Shipped an LLM-driven document detection pipeline that lifted classification accuracy by 15%.
  • Built a credit-card redaction model that’s processed 510k client documents under compliance constraints.
Pratilipi · Data Scientist · 2022–2024

Owned the recommender stack and most of the internal ML platform.

  • Personalized recommendations — +50% reads, 4× subscriptions.
  • Migrated ML workflows to Metaflow; cut deployment costs 70%.
  • Built a semantic-search API (Sentence Transformers + FastAPI) that powered content deals with Disney and Star.
  • Maintained the internal ML platform end-to-end — AWS, Jenkins, Terraform, Grafana, Metaflow.
Bobble.ai · AI Intern · 2022
  • Conversational-data pipelines over terabytes of Hindi macaronic text.
  • Intent recognition and brand-sentiment models.

Now

Building MCP servers for systems I use every day — Telegram, Slack, Jira — and stitching them together with Claude. Writing more Go than Python for the first time in years. Reading more on retrieval, evals, and agent memory than usual.

Elsewhere