Skip to content
Mithun Das
Work

Two decades of software that had to stay up.

I’ve spent twenty years shipping products into production — designing them to scale, and to keep working when parts of them fail. The problems changed: from nationwide vehicle titling to document classification to AI platforms. The job didn’t. Make the hard technical calls, build the first version myself, and leave the team better at making the next one.

Built

Selected builds — and what each one taught me.

Vitu · 2025 — now

Document classification, automated end to end

The whole pipeline runs without hand-offs: CNN models trained on SageMaker, inference served across regions on ECS and SageMaker. It classifies hundreds of thousands of documents a day and is sized to absorb ten times that.

Lesson:Scale is a design decision you make on day one, not a fire you fight later.

Vitu · 2025 — now

Making AI a safe tool for a whole company

An enterprise AI gateway with per-person budgets, usage tracking, and guardrails, so engineers and product managers can use Claude Code and Codex freely, within clear limits. Alongside it, a spec-driven way of building — specs generated from Jira, kept in sync with Confluence — now used across production projects.

Lesson:Adoption follows trust. Governance is what makes people comfortable moving fast.

Cox Automotive · 2015 — 2025

One rules engine for fifty states

Vehicle titling and registration runs on rules that differ by state and change constantly. Updating them used to be a weeks-long release. I designed an engine where they’re configured live — and where the UI renders itself from the rules — so changes ship in minutes and fewer transactions get stuck half-done.

Lesson:The best systems let the business change without asking engineering.

Where

Where the work happened — the title, and the role behind it.

2025 — now

Vitu

Senior Principal Engineer

Building the AI systems engineering runs on — and helping the whole org adopt them safely.

Architecture for AI systems in production, and the platform that lets the rest of engineering build with AI safely.

  • Document classification pipeline, automated end to end: CNN models trained on SageMaker, inference on ECS and SageMaker across regions — hundreds of thousands of documents a day, with 10× headroom by design.
  • Enterprise AI gateway on AWS Bedrock: virtual keys, per-user budgets, usage tracking and guardrails — governed access to Claude Code and Codex for engineers and product managers.
  • AI-native SDLC: Claude Code skills that turn Jira tickets into specs and keep Confluence in sync, now in use across production projects.
  • A shared skill repository, so teams improve AI workflows together instead of each reinventing them.

2015 — 2025

Cox Automotive

Senior Principal Engineer

Architected the platforms behind vehicle titling and registration, and took them serverless on AWS.

A decade building the systems behind vehicle titling and registration nationwide — and pushing the organization toward cloud, serverless, and applied AI early.

  • Architected registration and titling platforms for nationwide operations, where correctness isn’t optional.
  • Designed a live-configurable rules engine across all states — policy updates went from weeks to minutes.
  • Cloud-native, serverless-first architecture on AWS, adopted across multiple teams.
  • Explored edge AI (TinyML), IoT, and distributed ledgers years before GenAI went mainstream, and ran the intern innovation program two years running.

2024 — now

Appstack Solutions

Co-founder · Co-creator of Officebot

Building AI-native products of my own, end to end.

My own company. Product, architecture, infrastructure, billing — all of it.

  • Officebot: an agentic AI teammate in Slack, rebuilt across three eras of AI — from regex to NLP to agents.
  • Notelyt (in progress): an AI study companion — my daughter’s idea, which we’re building together.