$ they called me invalid_
I'm Madhan Kumar Reddy, an AI Infrastructure Operator in Singapore. I build private, on-premises AI systems that never send your data to the cloud: for firms that need AI but can't hand sensitive data to OpenAI or Anthropic.
No code hand-written: I spec, direct LLMs, and verify. Everything below is public, operating daily, and documented. Benchmarks, failures, and all.
Work with me Email me New here? Plain-English version
8 projects, documented in public 268 tok/s local LLM, measured 2 real quantum computers 0 manufactured wins
A local-first quantum machine-learning lab with one rule: a quantum number is never shown alone. Eleven experiments, every one raced against a classical baseline, two runs on real quantum hardware, zero manufactured wins.
Local image or text brief → a validated, engine-ready or marketplace-ready 3D asset on one RTX 5090: TRELLIS.2 / Hunyuan3D generation, headless-Blender mesh-ops (retopo, UV, PBR bake, real-world rescale), and an automated compliance gate that refuses assets failing a marketplace's published spec.
A fully air-gapped AI code-security auditor (cross-file taint + SAST + an LLM, hybridized) that finds vulnerabilities across files and returns deterministic, proof-carrying findings.
Self-hosted multimodal LLM infrastructure where 'client data never touches a cloud API' is enforced by routing, not policy.
Local AI video production: generation, object removal, masked edits, and 4K60 finishing, with a forensic-analysis bridge that starves diffusion models of room to hallucinate.
Fully local voice cloning, TTS, lip-sync avatars, music generation, and auto-dubbing: job-queue architecture, every model MIT/Apache-2.0.
A comic page goes in; a dramatized video comes out: panel detection, cloned character voices, narrator on captions, 2.5D parallax camera.
A Studio Ghibli-inspired productivity suite in Flutter: tasks, kanban, calendar, notes, pomodoro, and a physics tower-stacking break game with online leaderboards.
The self-hosting numbers people actually search for, measured on one RTX 5090: