Agent Harness Build & Audit
Your agent loops or calls the wrong tool. We trace the failure and fix context, tool use and recovery.
- An audited or implemented agent harness
- A failure map with reproducible traces
Asthra AI Labs
Make your model
work in your domain.
We help AI teams improve domain-specific LLM performance through evaluation, post-training, synthetic data and harness engineering.
For teams with an AI product
and a performance problem.
Fine-tuning made things worse. Your agent fails in production. Your evals cannot explain why. That is where we start.
Three ways to work with us
A clear scope.
A working system at the end.
Start with the bottleneck. Each engagement leaves your team with code, evidence and a practical next step.
LLM Performance Diagnostic & Evals
You cannot tell whether the model is getting better. We build task-specific evals and identify what is holding it back.
- A runnable evaluation harness
- A baseline report with failure breakdowns
Domain Model Improvement Sprint
Your model falls short on domain tasks. We curate data, run post-training experiments and measure the change.
- A versioned dataset and data-quality checks
- A reproducible training pipeline
Sprint or ongoing Model Improvement Partner.
The working method
Every change needs
evidence.
A completed training run is a starting point. The real question is whether your model performs better on the tasks you need it to do.
We start with
Your model, a task objective, and examples of where it fails.
Establish what good means.
Turn representative tasks and production failures into a held-out evaluation set. Agree on quality, latency and cost constraints before changing the system.
A baseline you can reproduce
Find the cause. Test the change.
Inspect outputs, traces and data. Isolate the failure, then compare a targeted change to the baseline: a better dataset, a training recipe or a harness fix.
Controlled experiments and a failure breakdown
Keep the gain. Catch the regression.
Re-evaluate on held-out tasks and inspect what got worse. Hand over the code, experiment record and regression checks so your team can keep improving.
A measured decision and a repeatable pipeline
Hard problems. Concrete work.
Selected experience and how we debug.
Chipforge AI / Domain training
Training looks healthy. The benchmark gets worse.
Investigating CPT and SFT regressions in an RTL model with reproducible VerilogEval runs and controlled experiments.
Debugging approach / Agent reliability
The agent worked yesterday. What changed?
Replay failing tasks, locate the first wrong decision, and separate model failures from context, retrieval and tool failures.
Founder-led engineering
Work directly with
the person doing the work.
I’m Avinash Mynampati. I work across domain-model training, evaluation and the systems around them. At Asthra, I stay close to the data, code and experiments with your team.
My work includes RTL model adaptation and research into how sparse MoE models route different tasks. That same attention to model behaviour shapes each engagement.
What is your model getting wrong?
Send Avinash a failing example and what you have tried.