Offline multi-agent diagnostic system for ADTC 2026 — llama.cpp, local RAG, HITL
# Africa Deep Tech Agent
**Africa Deep Tech Challenge 2026 — Laptop LLM Challenge**
**Domain:** Autonomous AI Agents
Fully offline multi-agent diagnostic system: **llama.cpp + GGUF**, local RAG, human-in-the-loop.
Measured peak RAM ≈ **2.1 GB** (well under 8 GB).
---
## Start here (submission)
| Document | Purpose |
|----------|---------|
| **REPORT.md** | Main challenge report |
| **JUDGE_ONE_PAGER.md** | 1-page judge summary |
| **DEMO.md** | 2-minute demo script |
| **SUBMISSION_CHECKLIST.md** | Pre-submit checklist |
| **FINAL_REVIEW.md** | Strengths / freeze advice |
| **ARCHITECTURE.md** | Deep technical design |
---
## Constraints
| Requirement | Status |
|-------------|--------|
| 100% offline at inference | ✅ |
| llama.cpp + GGUF only | ✅ |
| < 8 GB RAM (target < 7 GB) | ✅ ~2.1 GB peak |
| Multi-agent (Triage, Diagnostic, Resolution) | ✅ |
| Local memory / RAG | ✅ |
| HITL before actions | ✅ |
---
## Quick start
```bash
cd africa-deep-tech-agent
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
# Mock (no model file)
PYTHONPATH=. python -m src.main --backend mock --no-hitl "test"
PYTHONPATH=. pytest tests/ -q
```
### Real GGUF
```bash
pip install llama-cpp-python
python scripts/download_model.py # network once
PYTHONPATH=. python scripts/ingest_knowledge.py
PYTHONPATH=. python scripts/test_model.py
# Judge demo (CLI)
PYTHONPATH=. python scripts/demo.py --real --auto dry_run
```
Model: **Qwen2.5-1.5B-Instruct Q4_K_M** → `models/` (gitignored). See `models/README.md`.
### Desktop app (Tauri)
The GUI lives in `desktop/`. It talks to the local Python API (`src/server.py`) on `127.0.0.1:8000`. The window starts the backend automatically.
```bash
# One-click (macOS): double-click Africa Deep Tech Agent on the Desktop
# or:
bash scripts/launch_desktop.sh
# From source:
PYTHONPATH=. python src/server.py # terminal 1
cd desktop && npm install && npm run tauri dev # terminal 2
```
In the app …