mictopraise/smartie

Smartie: Practical business AI that works offline on the laptops Africa already owns. # Smartie Smartie is a sharply scoped, fully offline AI assistant for African micro and small businesses. It converts rough business facts into customer replies, marketing copy, payment follow-ups, and short operating plans on an ordinary CPU laptop. The competition artifact is the model repository itself. Smartie conforms to the official ADTC 2026 template: GGUF weights, `llama.cpp`, exactly two Corporate / Enterprise prompts, an idempotent public download script, and no network dependency during inference. **Gate 1 status:** Submitted on Devpost. The repository is now preserved as the Gate 1 submission snapshot. ## Project links - Devpost submission - Video demo - Technical report - Official profiler output ## Model choice The baseline is **Qwen2.5-1.5B-Instruct Q4_K_M** (Apache-2.0). Its approximately 1.12 GB weight file leaves substantial headroom below the 7 GB measured memory budget while retaining much stronger instruction following than ultra-small demonstration models. Q4_K_M is the initial quality/throughput compromise; it is a baseline to measure, not an invented performance claim. ## Quick start (Ubuntu 22.04 / WSL) Requirements: Python 3.11+, `curl`, and `llama-cli` from llama.cpp on `PATH`. ```bash bash download_model.sh python3 validate_submission.py python3 smartie.py payment "A customer owes NGN 32,500, due yesterday. Write a warm WhatsApp reminder under 55 words." ``` If `llama-cli` is elsewhere: ```bash LLAMA_CLI=/path/to/llama-cli python3 smartie.py plan "I run a small bakery..." ``` Smartie performs no network calls. Only `download_model.sh` uses the network, before inference. ## Tasks - `reply`: concise service and customer-care messages - `market`: channel-specific, low-cost marketing copy - `plan`: short operational plans grounded in supplied facts - `payment`: relationship-preserving payment follow-ups Use `python3 smartie.py --help` for runtime options. Defaults deliberately cap context at 2,048 tokens, generation at 220 tok …