🌐 theinvalid.me — portfolio, resume, and the failure log (new post weekly).
AI infrastructure engineer — I spec, direct LLMs, and verify. I build and operate sovereign, self-hosted AI systems: the planning, architecture, debugging direction, and operations are mine; the code is written by LLMs under my direction. B.E. in Computer Science. Open to freelance and full-time work.
One RTX 5090 workstation running three integrated stacks — every design decision, failure, and fix documented publicly:
Multimodal LLM serving (vLLM + Nemotron NVFP4, 275 tok/s, 180K context on 32GB) with routing-enforced privacy tiers — three fallback chains where "this data never touches a cloud API" is guaranteed by architecture, not policy. Includes the full Decision Log: why TensorRT-LLM lost six hours and vLLM won.
Local video production: FLUX.2 / Wan 2.2 / HunyuanVideo generation, VOID object removal (takes the object and its reflection), SAM3+VACE masked editing, RIFE 60fps + SeedVR2 4K finishing. Driven by a forensic-analysis bridge so diffusion models obey specifications instead of hallucinating.
Fully local voice cloning, TTS, lip-sync avatars (3 quality tiers), music generation, and auto-dubbing — job-queue architecture, all MIT/Apache-2.0 models, 25+ documented install lessons.
A comic page goes in; a dramatized video comes out — panel detection, cloned character voices acting the dialogue, narrator on captions, Ken Burns + 2.5D parallax camera. Built entirely on the three stacks above: the proof the infrastructure ships products.
🍃 komorebi
A Studio Ghibli-inspired productivity suite in Flutter — tasks, kanban, calendar, notes, pomodoro, and a physics tower-stacking break game with online leaderboards. Local-first SQLite; spec-driven with a documented upgrade path per feature.
discuss (multi-LLM) → consolidate → spec → execute (LLM-directed) → verify (gates) → operate
- Spec-driven: every phase has a written spec and a verification gate; snapshots before changes
- Honest by default: benchmark tables ship with empty cells and fill in public, weekly; failures are documented next to successes
- Sovereignty-first: local-first routing, Tailscale-only exposure, default-deny firewall, monitored backups
- 📊 Publishing weekly benchmarks across all three stacks
- ✍️ Writing up the build: the TRT-LLM→vLLM pivot, the SAM3 video-mode fix, the audio stack's 25 lessons
- 🌐 Portfolio site + case studies — coming soon
📫 Reach me: GitHub issues on any repo, or the contact on my profile.