Systems & AI Engineer
I build low-level stuff and focus on the product. Recent work spans a GPS navigation system deployed on a naval ship (now in patent filing), a custom Linux OS for embedded hardware, and an async AI agent SDK with 7k+ downloads.
Oct 2025 - Present
GOVVENS Pvt Ltd • Startup
Built Crowd Flow, a cross-platform desktop app that simulates how crowds move through a venue. A Tauri and Rust shell drives a Python simulation engine over JSON-RPC, validated by a 35-test suite tied to pedestrian-dynamics standards. I drove the build through an agentic Claude Code workflow and owned the architecture, the system design and the 13-screen UX, turning dense simulation output into a simple 0 to 100 risk score.
Apr 2023 - Jun 2025
Automa • R&D Lab
Co-founded and ran an independent R&D lab across about 7 AI, computer-vision and systems products. Led the architecture and engineering end to end. Standouts: NOUAR (naval GPS device, now in patent filing), the MoonLight agent SDK with 7k+ downloads, and a face-recognition pipeline running 200 streams on one GPU.
Feb 2025 - Jul 2025
Antares Life Cycle Solutions GmbH • Internship
Designed and specified the system architecture and data flow for an industrial telemetry platform (COD/BOD). Defined ingestion, processing, and visualization pipeline and built a minimal React (TypeScript) demo UI to validate the design and support client-facing technical discussions.
Jun 2023 - Sep 2023
SvaDhruthi (World Diabetes Forum | NGO) • Internship
Led a team of four on a health-information platform built to handle up to 1,000 concurrent users. Owned the backend architecture, a Flask and MongoDB system with a dual-agent RAG setup, a generator and a verifier, to keep answers grounded.
May 2026
A SaaS that personalizes cold email from uploaded product docs. LLM generation via my moonlight-ai SDK, grounded in LlamaParse briefs, with a CI-evaluated critic pass that lifted banned-pattern catch rate from 90% to 98%.
Jun 2026
A minimal async agent SDK with explicit control flow and no magic, so automations stay easy to debug. 7k+ downloads, now sunset at v0.3.0 and kept as a personal research bed.
Nov 2025
A Linux clipboard manager with image previews, emoji search and auto-paste, built because nothing on GNOME did it well. Rust with GTK4 and Unix-socket IPC.
Nov 2025
An evolution sim where agents learn to survive, on a neural network and genetic algorithm written from scratch in Rust and compiled to WebAssembly.
Oct 2025
A backtesting and execution engine over 10+ years of minute-level data. Pandas and NumPy with Numba JIT, real-time ingestion via the Kite Connect API, SQL storage. On hold until I find a real edge.
An independent R&D lab from my college years, with real budget and kit: an RTX 3090, a Jetson, a server and lab space. We pushed hardware and AI hard, built what people asked for, and dropped what did not. I led the architecture and the product calls.
Navigation Oriented Unmanned Astral Regulator
Patent PendingThe Navy needed shipboard electronics to sync to sunrise and sunset offline, on GPS alone. The first version was demoed on INS Investigator J15 and is now in patent filing, so I rewrote it as a pure Rust workspace with a Slint UI rendering straight to the screen, no compositor, on a custom Linux build for the Raspberry Pi 5. Holds within about 35 seconds of the USNO tables.
The agent platform that became the SDK
•Mar 2025
The first MoonLight, a full agent platform with a runner, recursive tool-calling, code execution, MCP and web search. The big agent bet did not hold, so I stripped it into today's SDK.
High-throughput face recognition
•Jul 2024
A face-recognition pipeline running ~200 streams in real time on one RTX 3090, load-balanced across 20 pods with self-retraining and hot-reload. Deployed for college attendance, then pulled after privacy pushback.
Low-level CV framework
•Oct 2024
A low-level C++ computer vision framework for face, object and depth recognition on Jetson-class devices. Shelved once academic work caught up and the hardware was slow to arrive.
Plain-language smart home control
•Nov 2024
A small Llama, fine-tuned and run locally, that controls a smart home from plain language. Worked well, but the local hardware cost never fit what people would pay.
Interview practice for students
•Sep 2024
An interview-practice tool with text quizzes and a live AI video mock-interview that returned a report. 200+ students used it before the credits ran out.
GPT inside WhatsApp
•Dec 2022
My first project, built the week ChatGPT launched. A middleware bridge that put GPT inside WhatsApp. Reached a couple hundred users before fizzling, but it earned the lab its early credibility.
Local fine-tunes optimized for consumer hardware via QLoRA.
Structured JSON output with custom fields. Trained to know when to use an action, why, and how to call it for HomeAssistant control.
General-purpose tool-calling model engineered to reliably output structured JSON payloads for an agentic framework.
Cross-domain sentiment classification. Placed 2nd for most innovative final-year project.
Predicts the next timeframe's HFT data. Trained on 15 years of minute-level market data split into 2-hour batches.
Built and published MoonLight, an async AI agent SDK with explicit control flow. Scaled to over 7,000 downloads before sunsetting at v0.3.0 as a personal research bed.
Built a production eval harness for an email critic agent. Proved deepseek-v4-pro achieves a 100% catch rate with zero false positives, and engineered a deterministic pre-pass to bring the flash model up to spec.
Recovered a dead demo on a naval ship in 15 minutes, tracing it to an HDMI and relay short and falling back over VNC. The captain then asked me to define the MVP requirements.
Ranked 62nd of around 1,400 finalists, from 150k+ submissions, solving a stochastic optimization problem with custom heuristics.
My go-to Linux distro is Arch with Hyprland, dual-booted with Windows (unless you have 4TB SSD and 2 GPUS for KVM). A mandatory addition to any dev portfolio.
2021 - 2025
Systems & AI Engineer
I build low-level stuff and focus on the product. Recent work spans a GPS navigation system deployed on a naval ship (now in patent filing), a custom Linux OS for embedded hardware, and an async AI agent SDK with 7k+ downloads.
Oct 2025 - Present
GOVVENS Pvt Ltd • Startup
Built Crowd Flow, a cross-platform desktop app that simulates how crowds move through a venue. A Tauri and Rust shell drives a Python simulation engine over JSON-RPC, validated by a 35-test suite tied to pedestrian-dynamics standards. I drove the build through an agentic Claude Code workflow and owned the architecture, the system design and the 13-screen UX, turning dense simulation output into a simple 0 to 100 risk score.
Apr 2023 - Jun 2025
Automa • R&D Lab
Co-founded and ran an independent R&D lab across about 7 AI, computer-vision and systems products. Led the architecture and engineering end to end. Standouts: NOUAR (naval GPS device, now in patent filing), the MoonLight agent SDK with 7k+ downloads, and a face-recognition pipeline running 200 streams on one GPU.
Feb 2025 - Jul 2025
Antares Life Cycle Solutions GmbH • Internship
Designed and specified the system architecture and data flow for an industrial telemetry platform (COD/BOD). Defined ingestion, processing, and visualization pipeline and built a minimal React (TypeScript) demo UI to validate the design and support client-facing technical discussions.
Jun 2023 - Sep 2023
SvaDhruthi (World Diabetes Forum | NGO) • Internship
Led a team of four on a health-information platform built to handle up to 1,000 concurrent users. Owned the backend architecture, a Flask and MongoDB system with a dual-agent RAG setup, a generator and a verifier, to keep answers grounded.
May 2026
A SaaS that personalizes cold email from uploaded product docs. LLM generation via my moonlight-ai SDK, grounded in LlamaParse briefs, with a CI-evaluated critic pass that lifted banned-pattern catch rate from 90% to 98%.
Jun 2026
A minimal async agent SDK with explicit control flow and no magic, so automations stay easy to debug. 7k+ downloads, now sunset at v0.3.0 and kept as a personal research bed.
Nov 2025
A Linux clipboard manager with image previews, emoji search and auto-paste, built because nothing on GNOME did it well. Rust with GTK4 and Unix-socket IPC.
Nov 2025
An evolution sim where agents learn to survive, on a neural network and genetic algorithm written from scratch in Rust and compiled to WebAssembly.
Oct 2025
A backtesting and execution engine over 10+ years of minute-level data. Pandas and NumPy with Numba JIT, real-time ingestion via the Kite Connect API, SQL storage. On hold until I find a real edge.
An independent R&D lab from my college years, with real budget and kit: an RTX 3090, a Jetson, a server and lab space. We pushed hardware and AI hard, built what people asked for, and dropped what did not. I led the architecture and the product calls.
Navigation Oriented Unmanned Astral Regulator
Patent PendingThe Navy needed shipboard electronics to sync to sunrise and sunset offline, on GPS alone. The first version was demoed on INS Investigator J15 and is now in patent filing, so I rewrote it as a pure Rust workspace with a Slint UI rendering straight to the screen, no compositor, on a custom Linux build for the Raspberry Pi 5. Holds within about 35 seconds of the USNO tables.
The agent platform that became the SDK
•Mar 2025
The first MoonLight, a full agent platform with a runner, recursive tool-calling, code execution, MCP and web search. The big agent bet did not hold, so I stripped it into today's SDK.
High-throughput face recognition
•Jul 2024
A face-recognition pipeline running ~200 streams in real time on one RTX 3090, load-balanced across 20 pods with self-retraining and hot-reload. Deployed for college attendance, then pulled after privacy pushback.
Low-level CV framework
•Oct 2024
A low-level C++ computer vision framework for face, object and depth recognition on Jetson-class devices. Shelved once academic work caught up and the hardware was slow to arrive.
Plain-language smart home control
•Nov 2024
A small Llama, fine-tuned and run locally, that controls a smart home from plain language. Worked well, but the local hardware cost never fit what people would pay.
Interview practice for students
•Sep 2024
An interview-practice tool with text quizzes and a live AI video mock-interview that returned a report. 200+ students used it before the credits ran out.
GPT inside WhatsApp
•Dec 2022
My first project, built the week ChatGPT launched. A middleware bridge that put GPT inside WhatsApp. Reached a couple hundred users before fizzling, but it earned the lab its early credibility.
Local fine-tunes optimized for consumer hardware via QLoRA.
Structured JSON output with custom fields. Trained to know when to use an action, why, and how to call it for HomeAssistant control.
General-purpose tool-calling model engineered to reliably output structured JSON payloads for an agentic framework.
Cross-domain sentiment classification. Placed 2nd for most innovative final-year project.
Predicts the next timeframe's HFT data. Trained on 15 years of minute-level market data split into 2-hour batches.
Built and published MoonLight, an async AI agent SDK with explicit control flow. Scaled to over 7,000 downloads before sunsetting at v0.3.0 as a personal research bed.
Built a production eval harness for an email critic agent. Proved deepseek-v4-pro achieves a 100% catch rate with zero false positives, and engineered a deterministic pre-pass to bring the flash model up to spec.
Recovered a dead demo on a naval ship in 15 minutes, tracing it to an HDMI and relay short and falling back over VNC. The captain then asked me to define the MVP requirements.
Ranked 62nd of around 1,400 finalists, from 150k+ submissions, solving a stochastic optimization problem with custom heuristics.
My go-to Linux distro is Arch with Hyprland, dual-booted with Windows (unless you have 4TB SSD and 2 GPUS for KVM). A mandatory addition to any dev portfolio.
2021 - 2025