Faisal Fayaz Faisal-Fayaz

+--[ faisal@github ]----------------------------------+
|                                                     |
|  FAISAL FAYAZ                                       |
|  AI systems / dev tools / interactive worlds        |
|  mode: open source                                  |
|                                                     |
+-----------------------------------------------------+

Faisal Fayaz

I build software that is useful, inspectable, and a little unexpected.

19 public repos · 5 live demos · 7 merged upstream PRs · synced 2026-09-25

faisal@github:~$ whoami
AI/ML + full-stack builder

faisal@github:~$ cat interests.txt
developer tooling  ·  local-first agents  ·  real-time systems
computer vision    ·  browser graphics    ·  open source

faisal@github:~$ echo $BUILD_PHILOSOPHY
"Make it work. Make the trade-offs visible. Leave the repo better documented."

`./projects --featured`

Repository	Signal
Sidekick	Local-first terminal companion — chat, voice, and 17 tools on your hardware. Published as `sidekick-agent` (`sk`). `Python` `Textual` `Ollama`
Physics Lab	C++ / SFML teaching lab: 17 scenes, 24 lessons, quiz, live instruments that measure F=ma. `C++` `SFML`
AutoApply	Semantic job-form autofill across major ATS platforms, with confidence scoring and local-first profiles. `Python` `FastAPI` `Chrome MV3`
Repropack	Wrap a failing command and turn it into a redacted, review-ready reproduction report—locally. `TypeScript` `CLI` `Node.js`
LeafScan	Offline crop-disease image triage with on-device TensorFlow Lite inference. `React Native` `Expo` `TFLite`
System Design Simulator · live	A 3D flight simulator for distributed systems. `React` `Three.js` `TypeScript`
IONSTORM · play	Neon WebGPU arcade shooter with WebGL2 fallback, seeded challenges, and offline PWA support. `WebGPU` `JavaScript` `Supabase`
Wild Roads · play	Procedural 3D world with vehicle physics, wildlife AI, and adaptive browser performance. `Three.js` `GLSL` `TypeScript`

./projects --more

Repository	What I am exploring
llama-autotune	C++17 llama.cpp hardware auto-tuner with a live Dear ImGui dashboard.
Codebase X-Ray	Interactive architecture maps: dependency graph, churn×complexity hotspots, blast radius.
EcoMind	Voice capture, transcription, and semantic recall for private, searchable memories.
Signal · live	Self-hosted speech-to-text with Faster Whisper, timestamps, and persistent history.

`./map --systems`

                             +----------------------+
                             |   THINGS I BUILD     |
                             +----------+-----------+
                                        |
              +-------------------------+-------------------------+
              |                         |                         |
      +-------v-------+         +-------v-------+         +-------v-------+
      | DEV TOOLS     |         | APPLIED ML    |         | REAL-TIME     |
      |               |         |               |         | WORLDS        |
      | Sidekick      |         | LeafScan      |         | IONSTORM      |
      | AutoApply     |         | EcoMind       |         | Wild Roads    |
      | Repropack     |         | Signal        |         | SysDesign Sim |
      | Codebase X-Ray|         | llama-autotune|         | Physics Lab   |
      +---------------+         +---------------+         +---------------+

`open-source --upstream`

Open source is where I learn fastest: read the real implementation, reproduce the edge case, write the test, and send the fix back upstream.

Soup (MakazhanAlpamys/Soup · ~6.2k★) — fine-tune LLMs from one YAML. Seven PRs merged:

#592 merged 2026-09-12 — fix(export): require explicit AWQ calibration data instead of silently downloading a 214k-row default set.
#588 merged 2026-08-28 — fix(train): make FSDP QLoRA dtype-compatible so 4-bit storage matches the compute dtype before wrapping.
#466 merged 2026-08-22 — feat(data): add Ollama and vLLM provider sampling to soup data best-of-n.
#437 merged 2026-08-17 — test(stream): pin the non-LoRA meta-guard scope so the adapter postcondition cannot silently widen.
#435 merged 2026-08-17 — fix(stream): verify adapter materialization across PEFT 0.18/0.19 instead of trusting a zero return count.
#432 merged 2026-08-16 — feat: DeepSeek V4 Flash GRPO recipe, catalog guards, docs, changelog.
#422 merged 2026-08-16 — feat: Qwen3.5 4B continued-pretraining recipe with regression coverage.

+ code should invite inspection
+ limitations belong in the README
+ a small reproducible fix beats a vague feature request
+ documentation is part of the implementation

`cat stack.json`

{
  "languages": ["Python", "TypeScript", "JavaScript", "C++", "SQL"],
  "ai_ml": ["TensorFlow Lite", "Faster Whisper", "llama.cpp", "Ollama", "PEFT / QLoRA", "LLM evaluation"],
  "frontend": ["React", "React Native", "Three.js", "WebGPU", "Textual"],
  "backend": ["FastAPI", "Node.js", "PostgreSQL", "Supabase"],
  "systems": ["SFML", "MCP"],
  "tooling": ["Git", "Docker", "GitHub Actions", "Vite"]
}

`./connect`

If you are building open developer tools, local-first AI, or something strange in the browser, I would like to hear about it.

> discuss an idea   : open an issue in the relevant repository
> collaborate       : send a focused PR or proposal
> say hello         : linkedin.com/in/faisal-fayaz
> portfolio         : faisal-fayaz-portfolio.vercel.app

status: building in public

ship useful things // document the sharp edges // contribute upstream

Navigation Menu

Faisal Fayaz Faisal-Fayaz

Achievements