Pocket AI — Milestone 1

Portable Windows MVP: hardware detection, local llama.cpp process management, streaming chat and clean shutdown.

What this milestone contains

• WPF desktop app (PocketAI.exe) on .NET 10.
• CPU/RAM/GPU detection.
• Automatic CUDA-first / CPU fallback runtime selection.
• llama-server bound only to 127.0.0.1 with a random port and random API key.
• OpenAI-compatible streaming chat through /v1/chat/completions.
• Windows Job Object cleanup plus process-tree fallback shutdown.
• Reproducible third-party asset lock file.
• Automated download + SHA-256 verification of Windows llama.cpp CPU/CUDA runtimes.
• Automated download + SHA-256 verification of a small official Qwen GGUF model.
• Runtime smoke-test and actual PocketAI.exe headless smoke-test.
• Windows GitHub Actions pipeline that builds, tests and emits the portable ZIP.

Pinned smoke-test stack

• llama.cpp: b10941, commit 4a89937354190cef5a97baf8eeb17336105eb72d
• CPU runtime: Windows x64
• CUDA runtime: Windows x64 CUDA 12.4 + matching packaged CUDA DLLs
• Model: Qwen/Qwen3-0.6B-GGUF, Q8_0, stored locally as models/chat/model.gguf

See assets.lock.json for exact SHA-256 hashes.

Fastest path on a Windows 10/11 x64 machine

Install the .NET 10 SDK, open PowerShell in the repository, then run:

.\scripts\prepare-win-release.ps1

It will:

1. fetch and verify all pinned AI assets;
2. test the CPU llama-server directly;
3. publish self-contained PocketAI.exe;
4. run PocketAI.exe --self-test --force-cpu;
5. produce dist\PocketAI-Milestone1-win-x64.zip.

The target PC that runs the resulting portable build does not need a separately installed .NET Runtime.

Run individual stages

.\scripts\fetch-assets.ps1
.\scripts\verify-layout.ps1 -Strict
.\scripts\smoke-test-runtime.ps1
.\scripts\publish-win-x64.ps1
.\scripts\smoke-test-pocketai.ps1

Normal user launch

After successful publish:

dist\PocketAI\PocketAI.exe

The app detects the hardware, tries the CUDA build when NVIDIA hardware is detected, and falls back to the CPU build if CUDA startup fails.

Headless integration self-test

.\dist\PocketAI\PocketAI.exe `
  --self-test `
  --force-cpu `
  --self-test-output .\dist\PocketAI\smoke-test-pocketai.json

A successful run exits with code 0 and writes passed: true to the JSON report.

Privacy behavior in Milestone 1

• inference server listens only on loopback;
• no account;
• no telemetry SDK;
• no cloud inference API;
• llama-server receives --offline;
• the app intentionally ignores hidden reasoning fields and displays final content only.

Current scope

This is still Milestone 1. It does not yet include RAG, document parsing, persistent chats, encryption, benchmark-based model selection or VRAM-aware autotuning.

See RUNTIME_SETUP.md, BUILD_NOTES.md and THIRD_PARTY.md for implementation details.

Milestone 3

See MILESTONE3.md for PDF, vector RAG, optional web research and local image generation.