About Me
Prathamesh Shetye
Staff Software Engineer at Pinterest · Android & Platform
Who I Am
I’m a software engineer with over 16 years of experience building things that ship to millions of people. For the last 11+ years, Android has been my primary craft — from deep framework-level work on carrier devices early in my career, to leading platform-scale initiatives at companies like Pinterest, Atlassian, and TripAdvisor.
What I Work On
My focus is on Android platform engineering: the systems, tooling, and architecture that other engineers build on top of. That means design systems, modularization, dependency injection at scale, API design, and making large codebases easier to work in. I care a lot about developer experience — reducing friction so teams can move fast without breaking things.
Earlier in my career I worked across the full stack — from C/C++ on embedded RTOS systems and Android NDK, to enterprise applications with PLSQL on AIX/UNIX. That breadth gives me a different lens when designing mobile systems today.
Journey
| Company | Years |
|---|---|
| 2018 – present | |
| Atlassian | 2017 – 2018 |
| TripAdvisor | 2015 – 2017 |
| Samsung (Contracting via L&T) | 2014 – 2015 |
| L&T | 2012 – 2015 |
| Mastek | 2007 – 2012 |
Tech I Reach For
Kotlin · Jetpack Compose · Android SDK · Java · Design Systems · Dependency Injection · GraphQL · C / C++, Python, Typescript
Beyond Work
Outside of my professional work, I’m a dedicated homelab enthusiast, constantly tinkering with self-hosted infrastructure to understand the ‘how’ behind the ‘what.’ I’ve found AI tools to be a massive force multiplier here—accelerating my projects while deepening my understanding of complex service architectures. When I’m away from the terminal, I’m usually watching or playing soccer, or diving into the latest piece of emerging tech.
My Local AI Stack
Most of my AI work happens on machines I own, on my own network — because the point is to understand the system end to end, and to keep data at home when it matters. Three pieces make up the current setup:
- A local LLM server. A workstation-class Blackwell GPU (NVIDIA RTX PRO 6000, 96 GB) running vLLM, serving an open-weights model (Qwen) with a long context window. A thin proxy layer in front handles API-format translation and transparent local web search, and the whole thing is observable in Prometheus + Grafana. It’s there so my coding agents (Claude Code, opencode) can run locally, for real, not in a demo. Building it was a journey.
- A split-brain agent deployment. Two machines over Tailscale, each doing what it’s best at: a low-power mini PC runs the always-on agent gateway (OpenClaw) as a hardened, sandboxed appliance, while a GPU box serves local models through Ollama. Routine work runs on the local model; the hard reasoning falls back to a cloud model. Nothing is exposed to the internet. Here’s the deployment, end to end.
- Sandy — a sandboxable assistant. An assistant I built that talks to internal systems only through declared MCP servers and is designed to never leave the sandbox it runs in — zero egress by construction, not by policy. Reports come back with every claim footnoted to its source. It’s the part of this stack that started as a project and became the interesting thing to build. Read the post.
The thread tying these together is the same as the rest of my work: I’d rather understand, and be able to prove, what a system is doing than take the box on faith.