Agri price intelligence
Back to all work
Case study · AgriTech
A domain LLM trained for agriculture
Domain-adaptive pre-training and fine-tuning of a Gemma-family model for Hindi, Telugu and code-mixed agricultural text.
Headline result
~3B tokens
domain-adaptive pre-training
- Industry
- AgriTech
- Client
- TBD
- Result
- ~3B tokens
The problem
General-purpose models didn't know the domain, and they handled Hindi, Telugu and code-mixed text poorly.
What we built
We ran domain-adaptive pre-training of a Gemma-family model on around 3 billion tokens using H100 GPUs, then fine-tuned it with QLoRA.
The tokenizer was extended for Hindi, Telugu and code-mixed text. Tenant-isolated LoRA adapters sit behind a four-layer, cost-aware router.
Outcome
Quality is measured against a 1,000+ item, expert-validated trilingual benchmark.
Next case study
Read case study
Get in touch
Have a similar problem to solve?
Talk directly to a senior engineer about what you're building.