Agent audit
A one-week, fixed-price review of a LangGraph or LangChain agent that already answers real users. I read your production traces, find what's wrong and why, and leave you evals that keep catching it.
What I look at
I've fixed each of these in a production agent; the links go to the write-ups.
Answer quality
Wrong or ungrounded answers, wrong tool calls, and retrieval that pulls the wrong source, found in your real traces rather than demo prompts.
When the knowledge base disagrees with itselfEvals and tests
What your evals actually cover, what they miss, and which past failures nothing would catch if they came back.
Cost
Where model and tracing spend goes: duplicated state in traces, test traffic in production projects, oversized prompts, retries that quietly double calls.
Cutting LangSmith costs without going blindMemory and reliability
Memory that grows per conversation, clients built per request, missing timeouts, and failure paths that leave a user waiting.
The leak the garbage collector couldn't seeCode health
Types, tests and boundaries around the graph, so the next change doesn't silently break a node nobody is watching.
1,901 type errors to zero on a live AI codebase
What you get
- A written report: each finding with its root cause, the evidence from your traces, and a fix, ranked by impact.
- Three to five evals built from the top findings, running in your CI after I leave.
- Fixes for the top two findings as pull requests.
- A one-hour walkthrough with your team, recorded.
Price
$2,500
Fixed, for one agent. Half up front, half at handover.
First three teams: $1,500. In return, a short testimonial if you found it useful.
How it works
Step 1
A 30-minute call
Free. We check that your agent is in production and that an audit is the right tool.
Step 2
Read-only access
Your traces (LangSmith, Langfuse or similar) and the repository. I sign your NDA. No write access to production and no customer database.
Step 3
Five working days
I review the traces and code, reproduce the worst failures, and write the evals.
Step 4
Handover
Report, evals and pull requests, then the walkthrough.
Who it's for
Teams with an agent in front of customers, where someone is asking: are the answers right, why is the bill growing, why does memory keep climbing? If your agent isn't live yet, a design review fits better. Ask me about that instead.
Who I am
Tariq Amarneh, a senior software engineer in Amman, Jordan. I build and run production AI agents at BusinessChat, and before that I was at Amazon. I wrote python-leak-hunter, an open-source tool for finding memory leaks in Python services, and I contribute fixes to Langfuse.
Start with a call
Pick a time that suits you, and tell me what the agent does and what's worrying you.
Or email me at tariqs.naser@gmail.com. I reply within a day.