Evidence-grounded treatment reasoning for pharma and biotech research teams.
Answers are built step-by-step from retrieved tool results, with the evidence trail shown — not recalled from memory. Claims we cannot match to retrieved evidence are flagged.
Pramana is a research-assist agent that reasons iteratively over a curated library of 212 biomedical tools — including FDA-approved drug labels going back to 1939, Open Targets, ChEMBL, and EuropePMC — to answer drug and treatment questions. Instead of answering from parametric memory in one shot, it decides what evidence is missing at each step, calls the right tool, retrieves real evidence, and updates its analysis, then returns a synthesized, citation-grounded answer.
It is built on ATHENA-R1, an open, MIT-licensed research release (Qwen3-8B fine-tuned + RL-trained), so the reasoning is small, fast, and cheap to run without giving up rigor.
Accuracy across 3,168 tasks — ahead of GPT‑5 by 17.8 points and DeepSeek‑R1 (a 671B‑parameter model) by 25.9 points. An 8B model outperforming models many times its size, through tool-grounded reasoning rather than raw scale.
Accuracy across 456 real cases — ahead of GPT‑5 by 10.7 points.
Validated three ways beyond benchmarks: a blinded preference evaluation by experts from 28 rare-disease organizations, where Pramana was preferred over reference models on all 8 evaluation criteria; review of complex hospitalized cases by practicing physicians; and adverse-event hypotheses checked against 5.4 million real patient EHR records, where true safety signals showed adjusted odds ratios of 1.48–1.84 and known-negative controls correctly stayed null.
Every answer shows the tool calls and results it was built from, and flags claims it could not match to them — a direct answer to the hallucination problem that limits most clinical-LLM products.
You ask a question in plain language. Pramana plans which evidence sources it needs, queries them, and returns a grounded answer with its reasoning and sources shown, not just a final claim.