Repo-Specific Post-Training
We adapt open coding models to your internal APIs, architecture boundaries, and engineering conventions using synthetic repo tasks.
Outcome
Learns your patterns
Enterprise AI Reliability
Repo-Aligned Coding Agents For Complex Enterprise Repositories
Experts still have to catch confident agent errors
Noam Brown shared that coding agents helped him iterate faster, but they also made confident, repeated mistakes that required expert intervention to resolve.
In a poker-solver build, agent outputs looked plausible but were still wrong in key cases, showing how overconfidence becomes risky in specialized technical work.
Open sourceEnterprise teams are moving from copilots to agents, but generic agents break down on large private codebases. They miss hidden repo rules, generate plausible-but-wrong changes, and increase review and security burden. What's missing is a repo-aligned adaptation and verification layer that turns output into trusted, shippable changes.
Best fit
1,000+ engineer organizations
Codebase
Long-lived, high-coupling repositories
Deployment
VPC or on-prem control boundary
Solution
A three-layer system that adapts to your codebase, verifies every change, and deploys safely inside enterprise controls.
Continuous Verification Harness
Every output is evaluated against CI, tests, quality checks, and security controls so trust is based on evidence, not optimism.
Outcome
Proof attached to each change
Enterprise Deployment Controls
Deploy in customer VPC/on-prem with governance, audit logs, and phased rollout so adoption scales safely across teams.
Outcome
Governance built in
Evidence
Aryabhata 1.0 proves that our process isn't just a theory. It is a 7B parameter model built on these exact principles which successfully outperformed the world’s most advanced frontier models in elite mathematical reasoning. (Aryabhata 1.0)
Aryabhata 1.0 vs Frontier models
| Model | In-distribution | Out-of-distribution | |||
|---|---|---|---|---|---|
| JEE Main Jan 2025 Accuracy | JEE Main Apr 2025 Accuracy | Avg. Tokens per Response | MATH 500 | GSM8K | |
Aryabhata 1.0 | 86.0 | 90.2 | ~2K | 83.6 | 94.8 |
Gemini 2.5 Flash | 83.0 | 83.5 | ~1.5K | 93.6 | 85.1 |
GPT-4.1 | 75.0 | 80.0 | ~1.8K | 86.6 | 94.0 |
GPT-4o | 46.5 | 44.0 | <1K | 69.2 | 94.6 |
Bigger models do not always mean better, and prompting alone does not always work. Aryabhata 1.0 beat frontier models because we tuned the model for this exact use case.
Math to Code
The same process behind Aryabhata 1.0 can be applied to enterprise coding agents. We need to expose the agents to the workflow the same way we onboard human engineers and let the agents learn through feedback, not through prompts.
Let's talk!
Bring us your repositories and workflows and we can discuss how custom agents can solve your specific challenges.
Leadership
We are a team of passionate builders focused on making AI coding more reliable in real enterprise environments. Our goal is to improve efficacy of the models, and reduce the risk and friction of adoption for engineering teams.
Sachin has 10+ years of experience in training and build models using RL. Here are a few highlights over the decade:
- Trained Aryabhata 1.0, outperforming OpenAI and Google's frontier models on mathematical reasoning for JEE maths problems
- Built self-driving car agents at Huawei Research with real-world deployment
- Developed algorithmic trading agents at JP Morgan Chase
Previously at
Rohith is a deep learning researcher and engineer with a track record of shipping SOTA models to production. Here are some highlights:
- Created Table Transformer, achieving state-of-the-art results in table extraction and structure recognition
- Engineered self-improving pipelines for VRU and emergency vehicle detection in production autonomous systems
Previously at
Resources
Paper · 1/3
Aryabhata 1.0: RL Post-Training For Compact Reasoning
Our 7B parameter Aryabhata model outperforms OpenAI's O4 mini and Google's Gemini Flash 2.5 on mathematics benchmarks—designed to serve millions of students at scale.
Open paperFrom Rewards to Reasoning: A Talk By Sachin
Session on how reward signals and RL techniques are used to shape language-model reasoning behavior.
Open talkTraining Large Reasoning Models With RL and Search
Technical deep dive on experimental learning loops for reasoning model training and evaluation.
Open talkFAQ
Quick answers on security, deployment, and measurable outcomes.
How is customer code protected?
AthenaAgent is designed for private deployment in customer-controlled environments with strict access boundaries and auditable activity logs.
How do you prove reliability?
Every generated change runs through a verification harness that checks CI, tests, policy rules, and security tooling before it is trusted.
How does rollout work?
We start with a scoped workflow and expand gradually as measurable quality and cycle-time metrics demonstrate sustained gains.
What ROI should teams expect?
Target outcomes include lower review burden, faster issue resolution, fewer regressions, and clearer governance over agent behavior.
Can this fit existing engineering systems?
Yes. AthenaAgent is designed to integrate with existing CI/CD, security scanners, and review workflows rather than replacing them.