AI Engineer, LLM Applications

Two of our case studies are conversational products: a knowledge-base assistant and a support agent living inside a fintech app. Both needed grounding, guardrails and a way to tell whether an answer was any good. That's this job, and the demo is the easy part.

You'd sit inside a product team instead of a separate AI wing, shipping the feature, its evaluations and its monitoring together, and being there when the client asks why it said that.

Expect to spend as much time on retrieval quality, prompts under version control and failure handling as on anything that looks like machine learning in a conference talk.

Being AI-native is the whole company here, not one team. You'd be the person others come to when their agent misbehaves, and the person who says plainly when a feature doesn't need a model at all.

• Design retrieval that works. Chunking, embeddings, hybrid search, reranking, and the judgment to use plain search when it wins.
• Build agents with brakes. Tool calling with clear boundaries, human handoff, and behavior you can explain the next morning.
• Measure instead of vibing. Evaluation suites tracking quality, cost and latency per release, not per impression.
• Put guardrails where they matter. Prompt injection, data leakage, and answers the client couldn't defend to a regulator.
• Say when it's the wrong tool. Sometimes the answer is a search box and a cleaner database, and someone has to say so.

• Two Years Shipping Production Code: Python or TypeScript, with at least one LLM product real users touched.
• Retrieval Built Properly: chunking, embeddings, hybrid search and reranking, on pgvector, Qdrant or whatever the client can run.
• Agent Frameworks Without Religion: LangGraph, the vendor SDKs, or plain functions, picked for the problem rather than the changelog.
• Structured Output Discipline: JSON schemas, tool calling and validation, so the code downstream isn't parsing prose.
• Python or TypeScript to Production Standard: typed, tested, packaged, and deployable by someone who isn't you.
• Prompt Versioning: prompts in source control with evaluations attached, not pasted into a chat window and forgotten.
• Guardrails and Failure Modes: injection, leakage, refusals, and the fallback for when the model is having a bad afternoon.
• Hands-On With Models: OpenAI, Anthropic or open-weight, plus at least one vector store you have tuned.
• Evaluation Discipline: you measure before you tune, and you can show the eval that justified the change.
• Cost and Latency Aware: you know what your design costs per thousand requests, roughly, without opening a calculator.
• Client-Facing Clarity: you can discuss what the model can and can't be trusted with, without hype in either direction.
• Field-Aware, Not Faddish: you keep up without dragging every new paper into a client's production system.
• Fluent English, Written and Spoken: most of our work happens in a written thread or on a call, and both have to be clear. Arabic is a genuine advantage across the region.
• Builds With Agents Too: you use coding agents heavily in your own work, with the same skepticism you apply to the ones you ship.

• Rigorous: a demo that works once doesn't impress you.
• Curious: you read the failure cases before the success metrics.
• Grounded: you're interested in what helps the user, not what sounds impressive in a standup.

• Fine-tuning, distillation, or running open-weight models in production.
• Speech, document extraction or other multimodal work.
• Familiarity with data residency rules in the region.

• Remote across MENA. Work from wherever you are in the region, with hours built around your team's overlap rather than someone's attendance spreadsheet.
• Work that ships. Products go live in weeks, not quarters. When someone asks what you have been building, you can open it on your phone and show them.
• Small teams, no ticket queue. Four to eight people carry a product from the first client call to handover. Nobody hands you a spec and disappears.
• Kit and tools you choose. A budget for your setup, and for the editor, licenses and services that make you fast. No three-week procurement saga.
• A learning budget with time attached. Money for the course or the conference, plus room in the week to use it. The second half is the part most places skip.
• Client contact from day one. You sit in the room where the decisions get made. Plenty of engineers wait a decade for that seat.
• Time off people take, and no timesheets. We measure engagements, not humans in fifteen-minute increments. Book the holiday, go, come back.