Tsela — Offline University Admissions Guidance for Botswana

Tsela is an offline, on-device AI assistant that helps Botswana students navigate university admissions — points requirements, fees, and deadlines — with zero internet dependency at inference time. Built for the Africa Deep Tech Challenge 2026.

Full technical writeup, design decisions, and benchmark results: REPORT.md

Why

Reliable admissions information for Botswana's universities is scattered and often unreliable — and not every student researching their options has consistent internet access. Tsela runs entirely offline once set up, using a small quantized language model backed by a verified knowledge base covering six institutions: University of Botswana (UB), BIUST, BA ISAGO, Botho University, Limkokwing University of Creative Technology, and the Botswana School of Business Sciences (formerly BAC).

Setup

Requirements:

• Python 3.10+
• llama.cpp built with llama-cli available — either on your system PATH, or point to it via the LLAMA_CLI_PATH environment variable

1. Clone the repo:

git clone https://github.com/mulaxprime/Tsela.git
cd Tsela

2. Download the model weights:

bash download_model.sh

This pulls the GGUF model file (~1.1 GB) from a public Hugging Face repo into model/. Safe to re-run — it skips the download if the file already exists.

3. Point to your llama-cli binary (if it's not on your system PATH):

export LLAMA_CLI_PATH=/path/to/llama-cli

4. Run the GUI:

python app.py

Project Structure

Tsela/
├── app.py                 # Dark-theme tkinter GUI
├── run_model.py            # Model inference + points-lookup bypass logic
├── knowledge_base.py        # Verified university admissions data
├── download_model.sh        # Downloads model weights (required for ADTC submission)
├── metadata.json             # ADTC submission metadata
├── REPORT.md                  # Full technical writeup
├── benchmark.py                # Local latency/RAM/throughput benchmarking script
└── benchmark_results.csv        # Sample benchmark output

Known Limitations

See the "Known Limitations" section in REPORT.md for a full, honest accounting — including current gaps in hallucination protection outside points-lookup queries, incomplete data for two institutions, and English-only language support.

License

MIT — see LICENSE.