fablecut
Zero-dependency browser video editor that AI agents can drive. JSON timeline, MCP and REST control, live-reloading UI.
JavaScriptMCPVideo
ronak parmar
backend and full-stack developer/vadodara, india
I build compilers, developer tooling, and AI systems.
/work
deepseek-v4-in-c
DeepSeek-V4 inference in C99, streamed off NVMe. Runs the 284B Flash checkpoint from 3.2 GB of RAM at 1.81 s per token, verified against PyTorch to 2.9e-6.
C99InferenceNVMe
dragoon
A JavaScript compiler written from scratch in C. Full pipeline: lexer, parser, AST, and native code generation through a QBE backend.
CQBE
aict
CLI that emits structured XML and JSON for AI agents to consume directly, bridging traditional tooling with agent workflows.
GoMCPCLI
lapdeck
Turns a phone into a remote deck for a Windows laptop. App launcher, touchpad, keyboard, live screen view, media and power. One Node process, no cloud.
NodeJavaScriptWindows
foundation cli
Dependency-aware full-stack scaffolder. Composes Next.js and Node monorepos into a working, conflict-free app in under three minutes.
TypeScriptNodepnpm
burrow
Voice-driven AI research assistant on an infinite canvas. Local-first, bring your own key.
TypeScriptVoiceLocal-first
terrafirm
Interactive 3D globe that zooms into a live street map, layering OpenStreetMap and Wikidata geodata with H3 spatial indexing.
ReactWebGLMapLibre GLH3
road damage detection
Fine-tuned YOLOv8 model for road damage and pothole detection, built alongside the Vadodara road analytics platform.
PythonYOLOv8Vision
/experience
- 2026, now
technical lead
KesariX Technology, remote
- Lead technical direction and delivery of AI-driven products, owning architecture and hands-on development.
- 2025
project engineer, intern
Tinkering Hub, Parul University
- Built a web application for road-damage analytics and reporting across Vadodara.
- Co-developed an R-CNN model for pothole and road-damage detection.
- 2024
project intern
Tinkering Hub, Parul University
- Prototyped AI and Flask-based applications across a ten-month engineering internship.
/stack
languages
backend
vision and ml
frontend
data
tooling
/recognition
/me
I write software close to the metal, and tools that sit right next to the work. A JavaScript compiler in C with a QBE backend. DeepSeek-V4 inference in C99, streamed off NVMe, running a 284B checkpoint out of 3.2 GB of RAM. The interesting part is usually the constraint.
Lately most of what I build is meant to be driven by agents: a browser video editor with an MCP and REST surface, a CLI that speaks structured XML instead of prose, a research canvas that runs local-first on your own key. I keep landing on the same preferences. Zero dependencies, no cloud, nothing you cannot run yourself.
The rest is ordinary and useful. Vision models watching the roads in Vadodara, full-stack products, a scaffolder I built because I got tired of not having one. I learn by building the thing, and I would rather ship something rough that works than plan something perfect that does not.