Browse categories
Explore
Fiverr Pro
English
$
USD
I will build an llm evaluation harness regression tests and hallucination guards
Majhar
About this gig
If you cannot measure your AI feature, you are shipping on vibes.
I owned the AI systems layer on a live consumer credit product: prompts kept as versioned artefacts, responses constrained to schemas and validated at the boundary, grounding checks tracing claims back to source records, and a regression set run against every prompt and model change with outputs sampled and scored by hand.
WHAT I BUILD FOR YOU
An evaluation harness wired into your actual pipeline, not a notebook
A regression set built from your real traffic and your edge cases
Automated scoring: schema validity, groundedness, task success, refusals
CI integration, so no prompt or model change ships unmeasured
Grounding and citation checks that catch invented facts
Output schemas with validation, retries and fallback chains
A human-review sampling queue for what automation cannot score
Drift monitoring for when a provider updates a model under you
WHAT YOU GET
A baseline you can point at, and a number that moves when you change something.
This is engineering, not certification. I measure and harden. I cannot promise a model will never be wrong, and nobody honest will.
Get to know Majhar
Majhar
Senior Software Engineer
- FromBangladesh
- Member sinceJul 2017
- Last delivery2 years
Languages
Bengali, English
Senior Software Engineer with 7+ years of experience designing scalable systems and leading development teams. Strong background in FinTech platforms, secure transaction systems, and high-performance backend architecture. Skilled in PHP, Laravel, Node.js, NestJS, MySQL, PostgreSQL, MongoDB, AWS, and DigitalOcean. I build reliable APIs, SaaS backends, and production-ready cloud systems focused on performance, security, and long-term scalability.
Other AI Development Services I Offer
FAQ
Our prompts mostly work. Why bother?
Because mostly is not a number. The first thing an eval set usually surfaces is which of your prompts quietly degraded when the provider updated the model underneath you, months ago, and nobody noticed.
Can you guarantee no hallucinations?
No, and anyone promising that is selling you something. I constrain, ground, measure and alert. The rate drops and becomes visible. It does not reach zero.
Do you need our production data?
No. Anonymised or synthetic samples matching the shape of your real traffic are usually enough, and I would rather you did not send me live customer records.
