The model is offline. The phone is not.
Team Limitless · iQOO Hackathon 2026 · Chennai City Battle · Developer Tools
A developer away from their desk can read code but not work on it. GitHub's mobile app shows you a pull request but won't let you fix it. Every AI coding assistant is a cloud service — it needs signal, costs per token, and still can't run your code.
iQForge turns the iQOO 15 into a real development environment. Clone a repository, edit files in a real editor, and let a code model running on the Hexagon NPU write, explain, review and debug alongside you. Then commit with an AI-written message and push straight back to GitHub — all from the phone.
Inference never leaves the device: no API key, no per-token cost, no round-trip, and it keeps working when your signal doesn't. Code moves over normal HTTPS to GitHub, where it already lives, and never to a third-party AI vendor.
| Stage | On the phone · Hexagon NPU | On the laptop · Office Kit |
|---|---|---|
| Get | Clone, pull, open a PR, scaffold a project | — |
| Write | Generate and complete functions in a real editor | Large multi-file refactors |
| Review | Diff review with risk and security flags | Whole-repo context passes |
| Debug | Stack trace → root cause → fix applied to the file | Errors needing a live run |
| Test | Lint, static checks, JS/Python snippets | npm test, full suites |
| Ship | AI commit message → commit → push → PR | npm install, builds, dev servers |
The rule: the NPU handles anything that is reasoning over code. The laptop handles anything that needs an operating system.
| Hexagon NPU | Local code model — free, instant, works on bad signal |
| LPDDR5X / UFS 4.0 | Model, KV cache, editor and file tree resident together |
| 50MP camera | OCR a diff straight off a monitor or whiteboard |
| Haptic engine | A failed build or BUG finding is felt, not missed |
| Gyroscope / proximity | Tilt-scroll long diffs; face-down locks the session |
| Office Kit | The build-server bridge to the laptop tier |
codebase/
├── android-app/ the phone app — Kotlin, Jetpack Compose, llama.cpp JNI
├── laptop-bridge/ FastAPI server — model escalation + toolchain execution
├── sample-diffs/ curated diffs for demoing
├── docs/ NPU benchmarks, GitHub integration design
└── CONTRACT.md the interface everyone builds against — READ THIS FIRST
# 1. Read the contract first — it's short, and it's why three people can build in parallel
cat CONTRACT.md
# 2. Open the app in Android Studio (runs with StubReviewEngine, no model needed)
# android-app/
# 3. Run the laptop bridge
cd laptop-bridge && pip install -r requirements.txt && uvicorn server:app --host 0.0.0.0 --port 8000Note: the native build expects llama.cpp vendored at
android-app/app/src/main/cpp/llama.cpp/— it's gitignored (too large to commit), so clone it there before building with the CMake path enabled.
| Component | State |
|---|---|
| Laptop-bridge escalation (FastAPI + Ollama) | ✅ Built and tested against sample diffs |
| GitHub PR fetch + share-intent | ✅ Built |
| Camera OCR, sensors, haptics | ✅ Built |
| On-device NPU inference (llama.cpp JNI) | 🚧 Integration written, build verification pending |
| Toolchain execution over Office Kit | 🚧 Next up |
Full task breakdown in TASKS.md and the GitHub Issues on this repo.
Mukeshkumar M · app shell, voice trigger, escalation logic Delfi A · on-device inference Nambert Jones L · laptop bridge, GitHub integration, camera & sensors
Built for the iQOO Hackathon 2026 · Developer Tools track