Browse categories
Explore
Fiverr Pro
English
$
USD
I will audit and fix your broken ai or llm pipeline
Rodrigo Torres
About this gig
Your AI feature worked in testing and breaks with real users. Or it runs every day and you have started to suspect the output is wrong.
That is what I do.
TWO REAL CASES
In one project I found a parser assigning columns by absolute pixel position. On older documents it produced output that was well formed and completely false, for months, with no error anywhere.
In another, a sign-language model was not learning because the feature vector silently dropped the hand landmarks it was supposed to read, on every single frame. The CI had been failing 10 out of 10 runs since the first commit and nobody had noticed. I took it to 58 backend and 21 frontend tests passing, and lint from 59 errors to zero.
WHAT I DO
I read your code, reproduce the failure, tell you the root cause in plain language, and either fix it or hand you the exact steps.
The quiet failures are the expensive ones. A crash gets fixed the same day. A pipeline that runs green while producing wrong data can cost you months.
BEFORE YOU ORDER
Message me with what is broken and what you have already tried. If I do not think I can help, I will tell you.
Get to know Rodrigo Torres
Rodrigo Torres
AI Engineer: RAG, Agents and MCP
- FromPeru
- Member sinceJan 2025
- Avg. response time1 hour
Languages
English, Spanish
I build LLM systems that keep working after the demo: hybrid RAG with measured retrieval quality, LangGraph agents with explicit control flow, and MCP servers. Production LLM work across four shipped products on OpenAI, Anthropic and Gemini. Full-stack in TypeScript, React, Next.js, Python and FastAPI, plus FlutterFlow for mobile and n8n for automation. Founder of Árkos: 50+ projects for 45+ clients since 2020. My work is public and auditable at github.com/RodrigoFK06/PrecioVivo. Two Anthropic certifications (2026).
My Portfolio
Other AI Development Services I Offer
FAQ
I cannot share my whole repository. Can you still help?
Yes. Often the relevant part is a few files, plus a description of the failure and a sample of the wrong output. Tell me what you can share and I will tell you if it is enough before you order.
What if you cannot find the problem?
You still get the written report with what I ruled out and why, which narrows the search. But I will tell you before you order if I think the case is outside what I can diagnose.
Do you work with models you did not build?
That is the normal case. Most of this work is reading someone else's pipeline and finding the assumption that stopped being true.
