Akshit Kr NagpalAI agent reliability For SaaS teams shipping AI agents
Your AI agent works in the demo and breaks in front of customers. I find out why, fix it, and show you the before-and-after on your own cases.
No pitch. You'll leave knowing what's most likely breaking and whether I can help.
Sound familiar?
If one of these is costing you customers, time, or a launch date, that's where we start.
“Customers keep finding its mistakes.”
Wrong answers, wrong actions, edge cases. I trace each failure to the prompt, tool, retrieval, or code behind it and fix it.
“Every prompt change fixes one thing and breaks another.”
I turn failures from your traces and tickets into tests your team reruns on every change, so the same bug doesn't ship twice.
“A model upgrade is coming and nobody wants to own it.”
I run your real cases on both models and show you what changes before your customers see it.
“Nobody can say whether it's ready to ship.”
We agree on what ready means, I close the gaps that matter, and you get a written list of what's still open.
Pricing
Fixed prices, agreed in writing before any work starts.
$1,500
3 days, fixed price
I go through your traces, tickets, and code, find what's actually breaking, and tell you what to fix first.
The full fee counts toward a sprint if you continue.
From $6,000
2 weeks, fixed price
I fix the failures that matter most, inside your codebase, and prove each fix on the same cases.
From $1,500
per month, cancel anytime
Models and prompts keep changing. I rerun your tests before each change and tell you what broke.
How it works
Tell me what the agent does and what's going wrong. I'll tell you honestly whether I can help.
I find the root causes, build a starter test set from your real failures, and hand you a ranked fix plan.
Take the plan to your own team, or have me fix it in a sprint. There's no lock-in either way.
Questions
They can write fixes quickly. The hard part is deciding what a correct answer is for your customers, finding which failures actually matter, and proving a change didn't break something else. That's the part I do, and your team keeps the tests afterwards.
No. I work with your existing setup, whether that's Braintrust, Langfuse, or your own scripts. If something is missing, we'll choose it together. Accounts, test data, and deliverables stay yours.
Access to the relevant code and examples of the failures, plus one person who can review test cases and answer questions. Usually that's a few hours a week.
I work only with the access you give me, and I'm happy to sign your NDA. Customer data stays in your systems, and test cases can be anonymized.
Each sprint is scoped to one problem. If the work turns up something larger, I'll explain what's left and your options. Any further work is agreed separately, at a fixed price.
Akshit Kr Nagpal, a senior full-stack engineer who builds AI systems and developer tools. I work alone, so the person on the call is the person doing the work.
Free, 30 minutes
Pick a time and tell me what the agent does and what's going wrong. On the call, I'll tell you honestly whether I can fix it.