I build things that sit between hardware and software: sensors, boards and radios on one side, backend services and AI on the other. By day I work on AI and industrial IoT at Suraj Informatics in Mumbai. In my own time I build open-source tools and run a homelab that is partly operated by my own AI agents.
In plain words: a camera reads a number plate, a tag is scanned, a scale weighs a truck, and a gate opens on its own while someone watches it happen live on a screen. Making that whole chain reliable is the kind of problem I enjoy, from the wire to the dashboard. These days I'm also teaching it to run AI on tiny boards, with no cloud needed.
Most of what's here started as "I wish this existed" or "I want to run this myself", and I use nearly all of it day to day.
Ferrum ⭐Fleet control for Proxmox VE: one dashboard for every cluster and node, with backups, HA, alerting, a REST API and a built-in MCP server.
|
Run llama.cpp on the Rockchip NPU. A ggml backend for RK3588, RK3576, RK3566, RK3568 and RK3562, with a one-line installer.
|
|
A private AI assistant that runs on your phone: chat, voice, document Q&A and screen understanding, fully offline.
|
Audited SSH, SFTP and tunnels for AI assistants, without ever exposing passwords or keys to the model.
|
|
Open IP-camera firmware for RV1103/RV1106 boards: RTSP/ONVIF, recording and on-board NPU detection such as ANPR.
|
A fast, modern download manager: multi-connection downloads, safe resume and browser capture. GUI, TUI and CLI.
|
🤖 AI agents & MCP servers
| Project | What it does | Stack |
|---|---|---|
| proxmox-mcp | 27 tools to run Proxmox VE (VMs, containers, storage, snapshots) from any AI client | Go |
| beszel-mcp | Ask Claude about server health, metrics, Docker containers and alerts | Go |
| local-computer-mcp | Safe, root-scoped local file tools for AI agents, with no shell execution | Go |
| file-operation-mcp-server | File, text, checksum, ZIP and JSON/CSV tools behind one MCP interface | Java |
🔌 Edge AI & hardware
| Project | What it does | Stack |
|---|---|---|
| SayWho | Who spoke when: speaker diarization app and API on GPU, Rockchip NPU, Apple Silicon or CPU | Python |
| Sabhalekh | Offline meeting transcripts in English, Hindi and Hinglish with NVIDIA Nemotron | Python · Tauri |
| rkllama.cpp-rk3576 | llama.cpp build tuned for the RK3576 NPU | C++ |
| socmon · rkdash | btop-style monitors for Rockchip and Allwinner boards: CPU, GPU, NPU, thermals and which process uses each | Go |
| springboot-tcp-server/client | Spring Integration TCP patterns for IoT device communication | Java |
🏠 Self-hosted platforms
| Project | What it does | Stack |
|---|---|---|
| Repoman | Single-binary artifact repository for Maven, npm, PyPI, Docker, Go, APT and more. Works air-gapped | Go |
| Docveta | Document management with OCR on GPU, NPU or CPU, plus SSO, AI Q&A and an Android app | Go |
| Whittle | Parametric 3D CAD on OpenCascade for desktop, server or browser | TypeScript |
| Vision Call | End-to-end encrypted chat and video calls for your home or office network | TypeScript |
| Lekhamint | Private money manager: an offline-first Android app and web app on one Go binary | Kotlin · Go |
| Geo Tracker · Shortr · Ferry | A Google Timeline alternative, a URL shortener with analytics, and LocalSend-compatible file sharing | Go · TS |
📱 Android & desktop apps
| Project | What it does | Stack |
|---|---|---|
| OpenWrt Manager | Manage your OpenWrt router from your phone, with nothing to install on the router | Kotlin |
| Gotify+ | Modern Material 3 client for Gotify push notifications | Kotlin |
| Lumen Launcher · Lumen Remote | An ad-free Android TV home screen, and a phone remote that pairs in seconds | Java · Kotlin |
| WinSpeed | Internet speed meter and data-cap alerts in the Windows taskbar | C# |
| Samjho Bolo | Hindi-first English learning app for the browser and Android | TypeScript |
Browse everything with search and filters →
The place where I break things on purpose:
- Compute: 6 Proxmox hosts (a 4-node SBC cluster plus 2 mini PCs) running Docker and LXC
- AI: Mac mini M4 and Radxa ROCK boards (RK3588 / RK3576 NPU) for local LLMs, OCR and speech
- Storage & network: CM3588 NAS on TrueNAS, with OpenWrt routing, VPN and SSO
- Operations: Beszel, Gotify and Home Assistant, plus my own MCP servers so AI can help run it
- Squeezing more tokens per second out of RK3576 / RK3588 NPUs
- Agentic workflows with Spring AI and MCP, with humans in the loop
- Making my self-hosted apps easy for non-technical people to install
From field signals to intelligent systems ⚡