benewende-dev/baarali-edge

Offline enterprise assistant for 8 GB laptops — Africa Deep Tech Challenge 2026, Laptop LLM track (corporate_enterprise). Every figure measured with the official profiler. # Baarali Edge — an offline enterprise assistant for the laptops West Africa actually owns Submission to the **Africa Deep Tech Challenge 2026 — Laptop LLM track**, domain `corporate_enterprise`. A quantised language model that runs entirely on a commodity laptop — 8 GB of RAM, integrated graphics, no network — and serves as the reasoning core of an offline assistant for small companies and public bodies: read the organisation's own documents, answer with citations, keep every byte on the machine. **Weights:** huggingface.co · **Demo video:** youtu.be/YUJUGaO2qjs (82 s) > **Status: submitted.** Every number in this repository comes out of the official > `adtc-profiler`; nothing is estimated. Where a figure is arithmetic applied to a measurement > rather than a measurement itself — `S_perf`, the scenario columns in the comparison tables — it > is labelled as such at the point of use. --- ## Why offline is the requirement, not a feature In the 2023 World Bank Enterprise Survey, **80,4 % of Ivorian firms reported electrical outages**, and affected firms put the loss at **2 % of annual sales**. The detail that decides the architecture is simpler than any statistic: **the laptop has a battery, the fibre box does not.** When the power goes, a cloud assistant becomes a spinning wheel at the exact moment the work is urgent — while a model living in the laptop's own RAM keeps answering. Add to that the documents that cannot leave the company, and on-device inference stops being a feature and becomes the requirement. The full argument, with sources and with an explicit list of what we **do not** claim, is in `USE_CASE.md`. --- ## Target profile The ADTC Standard Laptop — Intel Core i5 (10th–12th gen) or Ryzen 5, **8 GB DDR4**, integrated graphics only, 256 GB SSD, Ubuntu 22.04. Representative price: $150–$250 refurbished. Peak memory budget: **7 GB**. Runtime: **llama.cpp**, GGUF weights, CPU only (`-ngl 0`). --- ## Repository layou …