Tech · case study
BISAG-N, summer 2025
A farmer assistant that runs on a cheap machine and answers in your language.
BISAG-N is the Ministry of Electronics and IT's geo-informatics institute in Gandhinagar. My summer there was a language model project: the Kisan Call Centre has millions of recorded farmer questions and answers, and the idea was an assistant that could answer the same questions, in the farmer's own language, without a GPU behind it.
I fine-tuned Llama 3.1 8B on two and a half million of those records and got it to answer in under three seconds on an eight-gigabyte CPU-only machine.
How it works
Clean the transcripts
Two and a half million Kisan Call Centre records, in a mix of Indian languages, into training pairs.
Fine-tune without a GPU budget
LoRA adapters on a 4-bit quantised Llama 3.1 8B, which is what made the training feasible and the model small enough to serve on CPU.
Detect the languageCheck
The question's language is detected first so the answer comes back in the same one.
Answer, with live data
A FastAPI service answers in under three seconds and pulls live crop prices and weather into the reply.
A screen to try it on
A Streamlit front end so the institute could ask it questions.
What it changed
- Perplexity of 1.6 on the held-out set.
- Multilingual answers in under three seconds with no GPU in the loop.
- Certified by the institute's project manager and its additional director.
Built with Llama 3.1 8B, LoRA with 4-bit quantisation, PyTorch and Hugging Face, FastAPI, Streamlit.