Skip to main content
O
Omar Trkzi
Projects
Speaking
Blog
About
Book Session

Practitioner speaker · Morocco

I speak about what happens after the AI demo works.

Technical talks for teams moving from a compelling prototype to a system they can explain, evaluate, and trust.

EvaluationObservabilityFailure recoveryDependable AI product engineering

Software and AI engineer at OCP Group

I have spoken at Devoxx Morocco and BlaBlaConf. My public examples come from products and teaching systems I built.

Based in
Morocco
Languages
English · French · Arabic/Darija
Available talk

When Green Traces Lie: Evaluating Agentic Systems Beyond the Demo

Designed for Engineers and technical leads shipping LLM features.

A practitioner talk about finding failures that successful traces hide. CIBI supplies the real product implementation; cfp-dvma26-eval is a separate public teaching system built with invented data and deterministic mocks.

What the audience leaves with

  • 01Design reproducible failure cases
  • 02Instrument decisions so bad outputs are explainable
  • 03Understand where LLM-as-judge helps and where it can fail

Available formats

20min
25minDefault
45min

The 25-minute version is prepared for AI Engineer Paris; the structure adapts to the room.

Evidence behind the talk

Built evidence, clearly classified.

The product case and teaching system answer different questions. The page keeps that boundary visible.

01

real product implementation

CIBI observability

A Gemini-based scan and purchase judgment instrumented directly with LangSmith.

  • LangSmith run creation and update instrumentation surrounds the model boundary.
  • Best-effort tracing cannot block the scan response.
  • Tracing exposed missing token-usage telemetry even while the run appeared healthy.
  • Receipt-image privacy and retention remain explicit design trade-offs.
Development rolloutVisit CIBI
02

Public executable teaching and reproduction system

Multi-agent evaluation reference

A reproducible reference system using invented support-ticket data and deterministic mocks by default.

  • Deliberate routing, grounding, policy, injection, judge, and cost failures.
  • Regression metrics, inspectable traces, and thirteen passing tests.
Inspect source

Previous speaking

Verified appearances.

Devoxx Morocco

2024

“UX, But Dev-Exclusive!”

BlaBlaConf

2024

Organizer kit

Bios, logistics, and direct contact.

Travel base
Travel base: Morocco
Pronunciation:
OH-mar TURK-zee

50-word biography

Omar Trkzi is a Morocco-based software and AI engineer who builds dependable digital products. His public work explores AI evaluation, observability, product engineering, GIS, and production web software. He speaks from implementation experience, turning quiet failure modes into practical techniques engineers and technical leaders can apply immediately at their events.

100-word biography

Omar Trkzi is a Morocco-based software and AI engineer focused on dependable products and the engineering work that begins after an AI demo succeeds. Across teams ranging from Oracle to startups, he has worked at the intersection of customer experience, business constraints, and technical delivery. His public projects explore AI evaluation, observability, GIS performance, and production web software. On stage, Omar uses concrete implementations such as CIBI and reproducible teaching systems to show how healthy AI workflows can still fail. His talks give engineers and technical leaders practical ways to design failure cases, inspect decisions, and evaluate outputs with confidence.

Full biography

Omar Trkzi is a software and AI engineer based in Morocco. He enjoys the full software delivery lifecycle: shaping a useful product, finding the technical seams that matter, and making the result understandable to the people who will operate it. His experience spans Oracle and product teams across the Middle East, Europe, and Canada. Publicly, Omar builds and writes about dependable AI systems, product engineering, GIS, observability, and production web software. His speaking turns those implementation lessons into practical sessions for engineers and technical leads, especially the quiet failure modes that appear after an AI demo looks successful. One available talk uses CIBI’s real LangSmith instrumentation alongside a separate, executable multi-agent teaching system to explain reproducible failure design, trace interpretation, and the limits of LLM-as-judge evaluation. Away from software, he enjoys calisthenics, running, weightlifting, and martial arts.

Invite me to speak

Planning a room for engineers who need more than a demo?

Share the event, audience, and format. I will respond directly with the best fit for the program.

Email meLinkedIn

Omar Trkzi

AI & software engineer building agent workflows, automation systems, and product-grade web experiences.

Quick Links

ProjectsSpeakingBlogAboutContact

Connect

© 2026 Omar Trkzi. All rights reserved.