$ whoami
Satyajit GhanaHead of Engineering, AI & 3D Perception
Head of Engineering · Inkers Technology · Bengaluru, India
I build deep-learning systems, 3D perception, and high-performance infra.
Head of Engineering at Inkers, where I work on industrial AI — 3D perception, deep learning, and the systems that ship it to production.
287 repos · 222 stars · 181 followers · stats →
ask anything about me — my work, projects, papers, patents. answers are grounded in this site's content.
🤖 agents: this site is machine-readable. start at /llms.txt, fetch any page as markdown by appending .md, query /api/* for JSON, or call the MCP endpoint. humans: it's all just the website.
3D Model Anything: an iPhone scan, an agent, and a photo it keeps comparing against
2026-10-24Dessa: timing a speech-only conversation model from its own demo tracks
2026-10-24Generative Trajectory Planning: a diffusion prior that proposes plans and learns from the ones search finds
2026-10-24Hy-Nomas: an 8.6B world model whose memory is a KV cache, and what it takes to fit a 4090
2026-10-24Kimi K3.1 Pro: the same 2.8T model, more RL, and what the −65% counts
2026-10-24
- LLM architectureA field guide to attention mechanisms
- Agents & harnessesHy Memory: a 3B-active memory writer that ties Opus 5, and four launch numbers that belong to the engine
- Inference & servingJev in the browser: N stops being a loop bound
- Training & RLA fix for softmax and cross-entropy: the target gradient that rounds to zero
- Image & video generationQwen-Image-2.1: the 7B is 28% of the pipeline, and the licence is not Apache-2.0
- Developer tools & infraBM25: the ranking function that refuses to die
Tokenizers v1 and Combinatorial BPE: a faster BPE, and the six ways to spell 'the'
Runs on a laptop CPUQwen-Image-2.1: the 7B is 28% of the pipeline, and the licence is not Apache-2.0
Interactive explanationsEmbeddingGemma 2: the encoders are optional, the text model is not
Runs on a phoneNVIDIA Model Optimizer: from mtq.quantize to a 4.95-bit DeepSeek-R1
Runs on a consumer GPUstable-diffusion.cpp: 52 model versions in one binary, and a quantiser that never touches a convolution
Runs on a laptop CPU
Vritti UI
React 19 · Next.js 16 · Tailwind v4Animated React components and blocks you copy-paste into your shadcn/ui project. 560+ components, 240+ blocks, a visual theme editor, and a suite of creative tools (shaders, generative art, WebGL textures) — crafted for design engineers.
torch-point-ops
PyTorch · CUDA · C++High-performance PyTorch library for 3D point-cloud operations, with custom CUDA kernels.
fabrik
TypeScript · React · Vercel AI SDK@fabrik-sdk/ui — a generative-UI SDK for any LLM. The model decides what to render, not just what to say.
Quantum Orbital Visualizer
Next.js · Three.js · Rust (WASM)Interactive 3D atomic electron orbitals, rendered as hundreds of thousands of particles from the exact Schrödinger solutions — with swappable JS / Rust-WASM / C++23-WASM compute backends.
BOTCHA
Next.js · TypeScript · RedisAn inverted CAPTCHA that keeps humans out. Traditional CAPTCHAs ask you to prove you're human; BOTCHA asks you to prove you're a bot.
My site's chatbot was stuffing 273k tokens into every message. I gave it tools instead.
2026-07-20Giving my own search the Contextual BM25 treatment
2026-07-20Contextual Retrieval, with a runnable repro and a browser playground
2026-07-17
arXiv digest — 8 papers · standout ★
2026-07-22AdaFlash: Adaptive Speculative Decoding via On-Policy Distilled Diffusion Drafters