Pramana Research
The most interesting part of any answer is how it was reached. We build AI that shows you, step by step and source by source.
Vision
General-purpose models are fluent, but in high-stakes work fluency is not enough. Professionals need models that know their domain, read the documents in front of them, and leave a trail a reviewer can follow.
We take the best open-weight models, specialise them one domain at a time on data whose correctness we can prove, and release them only when they beat what came before. Because they are open-weight, organisations can run them on their own infrastructure, with their own data.
Training data comes from authoritative records, so every answer is correct by construction and cites where it came from.
Models read the source instead of recalling from memory, and show the reasoning behind each answer so it can be checked in seconds.
Measured on cases the model has never seen, reported by task, with plain statements of what a result does not prove.
Open-weight models deploy on your own servers, so sensitive data never has to leave your organisation.
Products
An analyst copilot for equity research. Reads reported financials, extracts figures and computes growth and margins with the working shown.
See results and capabilities →A model fine-tuned on your own documents, evaluated against your own questions, and deployed on your own infrastructure.
Talk to us →from Sanskrit gaṇanā, calculation, and gaṇaka, the one who calculates
Gananka reads what public companies reported to the SEC and does the extraction and arithmetic analysts do every day, citing the figures it used. It is a tool for analysts, not a stock picker: no buy or sell calls, no price targets.
784 questions on 10 companies held out entirely from training. Every answer is checked against the figures the company filed.
Overall accuracy
91.5%
up from 61.9% (base model)
Year-over-year change
+59.1 pts
21.6% → 80.7%
Operating margin
+52.6 pts
45.6% → 98.2%
Accuracy by task
Overall
YoY change
Operating margin
Extraction
What this shows, and what it doesn't: the evaluation uses the same question types as training, on companies held out from training. It proves the model learned these calculations on new data, not that it is a better analyst in general. Broader evaluations come with v0.2.
Work with us
We build private models for research teams, outsourcing firms and fintech platforms. Your data stays with you, the model runs on your own infrastructure, and every release is measured against your own questions before it goes live.