Logo Lanfrica

kish-00/karatasi

Domaine:

natural language processing

Type de record:

software
Créateur:
kis
Hôte:
Offline AI Document Processor for Kenyan Government Forms — OCR, form understanding, Streamlit UI, Swahili support ## ADTC 2026 Submission This repository is the **application source** for our Africa Deep Tech Challenge 2026 ("Laptop LLM Challenge") submission, **SME Brief** — an offline, bilingual (French/English), citation-grounded RAG assistant for African SMEs' own business documents. - **Submission repo** (fork of the official template — holds `metadata.json`, `download_model.sh`, `REPORT.md`): `github.com` *(public)* - **Track / domain:** corporate_enterprise · small-business finance · languages FR + EN - **Claims:** African Use Case Bonus (data sovereignty / federated-by-default) + 8 GB budget-laptop - **Eval:** 50/50 gold bilingual QA; fully offline; ~1.5–2 GB RAM - **Devpost copy:** `docs/demo/SUBMISSION.md` · **What we customized:** `docs/MODEL_CUSTOMIZATION.md` · **Demo script:** `docs/demo/DEMO_SCRIPT.md` --- # SME Brief — Offline Local RAG for African SMEs **SME Brief** is an offline retrieval-augmented generation (RAG) system that answers business questions about a small company's own documents — in French or English — with cited sources. Built for the **Africa Deep Tech Challenge 2026** ("The Laptop LLM Challenge" — AI that runs on 8GB RAM laptops, fully offline). ``` "Combien de factures sont impayées ?" "What was invoice AT-2024-0007?" │ │ ▼ ▼ ┌─────────────────────────────────────────────────────┐ │ QueryRouter → SQL intents OR semantic RAG (LLM) │ │ over a single-file SQLite store (documents + vec) │ └─────────────────────────────────────────────────────┘ │ │ ▼ ▼ "3 factures" "8,120.00 USD" files: [facture_…, …] files: [invoice_AT-2024-0007.pdf] ``` --- ## Why SME Brief? Small businesses keep their books in invoices, receipts, contracts, and bank statements — almost always in their local languag …