Ali Ahmed Samir

Abu Dhabi, UAE · Senior AI Engineer and AI Delivery Owner · 6 years

Production computer vision and agentic AI for UAE government.

I make AI systems faster, steadier under load, and safe to run inside on-premise, security-restricted environments, on NVIDIA GPU, CPU-only and BPU edge hardware. Six years of production AI, four of them in computer vision, the last five at Tahaluf Al Emarat.

See the work Download the CV ali.samir.1744@gmail.com
UAE work visa to Feb 202830-day noticeEgyptian · English, Arabic+971 50 608 2836

Open to senior and lead AI roles in Abu Dhabi and the UAE.

Ali Ahmed Samir speaking at an event, in a light jacket and white shirt, seen between two people in the foreground

Four separate measurements

  • 15% → 1%

    false alerts after adding a vision-language model as second reviewer. Live traffic, Smart Cities.

  • ~1 → ~100

    concurrent requests per CPU-only node. Federal ID-verification platform, 15+ models.

  • 36 → 82 FPS

    multi-model inference throughput, +127%, after re-architecting the inference engine.

  • −20% failures

    fewer system failures after redesigning the end-to-end pipeline of a Smart City RTSP video platform.

Delivered forClients include ADNOC, Darb and Mawaqif, and other UAE government and enterprise clients, through Tahaluf Al Emarat.

LangGraph · multi-agent · HITL · VLM · confidence calibration · TensorRT · Triton · DeepStream · ONNX Runtime · PyTorch · FastAPI · Docker · MLOps · CI/CD · DVC · Azure · Rust · RAG (personal projects)

What I built, under which constraints, and what it measured.

Nov 2025 to now · Serving under load · Federal ID-verification platform (ICP)

From about one request at a time to about a hundred.

A government ID-check service built from 15+ AI models could serve one request at a time on each CPU-only node. I traced the cause to a MediaPipe/TFLite step and moved that path to ONNX Runtime. A single node now handles about a hundred requests in flight.

My part
Found the bottleneck and migrated that path to ONNX Runtime.
Team
Technical lead on a team of 5 or more engineers.
Result
~1 → ~100 concurrent requests per CPU-only node.
Not stated
Latency, request payload and load-test method.

Nov 2025 to now · Agents for the parks · Smart Cities, Abu Dhabi

Agents that watch the parks.

A LangGraph multi-agent system, working with vision models, monitors 5 Abu Dhabi public parks with 28 to 55 cameras each. Agents call tools, open and close cases automatically, and hand uncertain ones to an operator for review.

My part
Designed the agent system and deployed it to production.
Team
Senior engineer on a team of 5 or more engineers; I designed the agent system.
Result
In production across 5 parks, 28 to 55 cameras each. Operators see escalations by confidence threshold, and the escalation rate is tracked.
Not stated
An outcome metric for the agent layer.

How a detection becomes a case · simplified

  1. Cameras
    5 parks, 28 to 55 each
  2. Vision models
    Detect and track
  3. LangGraph agent
    Tool calling, case state
  4. Confidence threshold
    Auto-close, or escalate to an operator (HITL)

The operator escalation rate is tracked, so threshold changes can be judged against operator workload.

Jan 2023 to now · Inference, reliability and trust

Faster inference. Steadier streams. Fewer false alerts.

Separate pieces of work with separate measurements, so the numbers below are not one benchmark.

  • 36 → 82 FPS (+127%). Jan 2023 to Nov 2025. I re-architected the multi-model inference engine with optimized matrix operations and GPU acceleration.
  • 20% fewer system failures. Jan 2023 to Nov 2025. I redesigned the end-to-end pipeline of a Smart City platform processing RTSP video, with better fault tolerance in production.
  • 10% higher precision, 35% fewer false positives. Jan 2023 to Nov 2025. I calibrated model confidence through PyTorch retraining, threshold tuning and CVAT data validation, and aligned alert thresholds with operational SLAs.
  • 65% lower inference time, 99.84% character-level accuracy. Jan 2023 to Nov 2025. I built a lightweight license plate model from scratch and extended it from the UAE to the whole GCC.
  • False alerts 15% → 1%. Nov 2025 to now. I added a vision-language model as second reviewer in the Smart Cities solution, shipped to production and measured on live traffic.
My part
Re-architected the inference engine and pipeline; built the LPR model; added the VLM second reviewer; calibrated model confidence.
Team
Technical lead on a team of 5 or more engineers.
Result
+127% throughput, −20% system failures, 15% → 1% false alerts, −35% false positives. Each is a separate measurement.
Not stated
GPU model and batch size behind 36 → 82 FPS; what a false alert is counted per; LPR dataset and conditions.
BPU port
Nov 2025 to now. I led joint engineering with a global automotive OEM to port our AI system to BPU chips, and fixed a layer-conversion failure to the .hbm format to cut BPU inference time. Not stated: team size, duration and the size of the gain.
Also
A generative AI API for the Abu Dhabi Presidential Cup finals: an LLM writes the post text and image processing composites it onto pre-provided designs.

Growth and leadership

  1. 2021Data Engineering Team Lead

    Led and evaluated 7 junior data engineers.

  2. 2022AI Engineer

    Proof of concept lifted from 28% to 89%. Employee of the Month.

  3. Jan 2023Senior AI Engineer

    Inference engine, LPR model, calibration.

  4. Nov 2025Senior AI Engineer and AI Delivery Owner

    Accountable for delivered quality. I mentor 1 to 2 AI engineers, review code and set delivery standards.

Headline figures, and what they do not say.

Each row comes from my CV. The last column lists what the CV does not state, which is not the same as not measured.

Claims, where each comes from, and what the source does not state.
ClaimWhereNot stated
15% → 1% false alertsSmart Cities, live trafficWhat a false alert is counted per; the time window.
~1 → ~100 concurrent requestsID verification, per CPU-only nodeLatency, payload, test method.
36 → 82 FPS (+127%)Multi-model inference engineGPU, batch size, which models.
−20% system failuresRTSP pipeline redesignHow a failure is defined; the comparison window.
65% lower time, 99.84% accuracyLicense plate model, character levelDataset and capture conditions.
+10% precision, −35% false positivesPyTorch retraining, threshold tuning, CVAT data validationBaseline model, dataset, threshold method.
28% → 89% on a proof of conceptNew algorithm from cross-team work; Employee of the MonthThe metric that was measured, and how.
97.8% mAP@50, 97% mAP@50-95License plate detection as 4-point polygonsDataset and test conditions.
5 parks, 28 to 55 cameras eachAgentic LangGraph systemAn outcome metric for the agent layer.

Ask the work, not the résumé.

Pick a question to see an example of how an answer would cite its sources. A live version would check that each citation resolves and every number appears in a source.

Example answers, written by hand

I designed a multi-agent LangGraph system with computer vision that monitors 5 Abu Dhabi public parks, 28 to 55 cameras each. It uses tool calling, opens and closes cases automatically, escalates by confidence threshold to a human operator, and tracks the operator escalation rate. Separately, I added a vision-language model as second reviewer in the Smart Cities solution, which cut false alerts from 15% to 1% on live traffic.
Sources: Smart Parks agent, VLM second reviewerIllustrative check, not run: 2 citations resolve; 5, 28, 55, 15% and 1% found in sources

Example answers, written by hand. They are not generated live and no check was run. A live agent would verify citations and numbers, not wording. To ask me something, use the email below.

The client systems aren't public. These are.

Rust · gRPC · 38 stars
Vision.rs

Object detection in Rust with a gRPC API for remote inference.
Python
YOLOv8-nms

Fuses non-maximum suppression into the YOLOv8 graph.
Rust
open-multi-agent-rs

Rust port of a multi-agent LLM orchestration framework.

All 10 posts · Deep Diaries on Substack

Credentials

BSc Computer Science
Kafrelsheikh University, 2020. First place, graduation projects competition.
Peer-reviewed paper
Evolutionary algorithm-based CNN for predicting heart diseases. Computers and Industrial Engineering, 2021
Certifications
Udacity: Agentic AI; AI Agents with LangChain and LangGraph (2025). LinkedIn Learning: MCP with agentic AI (2025). AWS ML Specialty, 2021 (expired 2024).

Contact

Hiring for a senior or lead AI role?

I am open to senior and lead AI roles in Abu Dhabi and the UAE. Or bring a hard problem: production vision, agentic systems, inference that has to hold. Describe it in a line and I'll reply with how I would approach it.

ali.samir.1744@gmail.com +971 50 608 2836 CV (PDF) LinkedIn GitHub Substack

Abu Dhabi · UAE work visa to Feb 2028 · 30-day notice