Contact

Frontier models know the world.
Savante models know your world.

We train custom specialist models for your field, built from your data, your evaluations, and your experts.

A general model is sufficient everywhere and exceptional nowhere in particular. We train an open model against the work your experts actually judge, until it becomes a specialist.

Evidence

Our work with PhysicsWallah: a 7B open model, RL post-trained for JEE mathematics, beats frontier models.

JEE Main April 2025 · pass@1 accuracy
Aryabhata 1.0
90.2%
o4-mini
85.8%
Gemini 2.5 Flash
83.6%
Qwen2.5-Math-7B
54.7%
JEE Main April 2025 pass at one accuracy
Model Accuracy
Aryabhata 1.0 90.2%
o4-mini 85.8%
Gemini 2.5 Flash 83.6%
Qwen2.5-Math-7B 54.7%

Qwen2.5-Math-7B is the open model Aryabhata was post-trained from: 54.7% to 90.2% on the same exam. Presented at the NeurIPS 2025 MATH-AI workshop. Read the paper →   The model →

Aryabhata was one domain. We train specialist models for yours.

Founders

Sachin Dharashivkar
CEO
  • RL post-trained Aryabhata 1.0 with PhysicsWallah
  • Led RL efforts for equity trading at JP Morgan Chase
  • Trained a model to play the collaborative game Overcooked at Unity
Aryabhata 1.0 paper → Training Reasoning Models, Part 1 → Training Reasoning Models, Part 2 → LinkedIn → X →
Rohith Pesala
CTO
  • Coauthored the Table Transformer and PubTables-1M at Microsoft Research (CVPR 2022)
  • Table Transformer models are downloaded several million times a month
PubTables-1M paper (CVPR 2022) → Table Transformer code → Detection model → Structure model →

If your models need to be experts, not generalists, let's talk.

sachin@athenaagent.com