Akram - AI Engineer
AI

Hello, I'm

Akram

|

AI Engineer specializing in Computer Vision and Generative AI. I build intelligent systems that perceive, reason, and act - from real-time surveillance pipelines to LLM-powered RAG assistants.

View Projects Contact Me
10+ AI Projects
2 AI Domains
8+ Frameworks

Domains of Expertise

Two main pillars forming my AI engineering practice

Generative AI

Building LLM-powered applications with local inference, RAG pipelines, and automated prompt optimization. Focus on practical, production-ready deployments.

  • RAG systems (FAISS + sentence-transformers)
  • Prompt engineering & optimization loops
  • Local LLM inference (Ollama)
  • Vision-Language Models (Qwen2-VL)

Computer Vision

Real-time video analytics for security & surveillance. Expert in YOLO-based detection, multi-object tracking, pose estimation, and anomaly detection pipelines.

  • YOLO object detection & pose estimation
  • Multi-object tracking (ByteTrack, Deep SORT)
  • Behavioral anomaly detection
  • Real-time RTSP/webcam pipelines

MLOps & Deployment

Shipping production-grade AI systems with robust alert mechanisms, REST APIs, and containerized deployments. Strong focus on reliability and real-world performance.

  • FastAPI & Flask REST backends
  • Docker containerization
  • Webhook & SSE alerting systems
  • ONNX runtime optimization

Technical Skills

Technologies I work with on a daily basis

{ } Languages

Python Bash HTML/CSS JavaScript YAML

⚡ AI / ML Frameworks

PyTorch Ultralytics YOLO Transformers (HF) ONNX Runtime InsightFace sentence-transformers PyTorchVideo

📷 Computer Vision

OpenCV Pillow Shapely Optical Flow Pose Estimation ByteTrack

🤖 Generative AI / LLM

Ollama Qwen2-VL FAISS RAG pipelines Prompt Engineering LLM-as-judge

🌐 Web & APIs

FastAPI Flask Streamlit Server-Sent Events REST APIs Webhooks

🔧 Tools & DevOps

Docker Git NumPy SciPy Pydantic argparse

Featured Projects

Production-ready AI systems across two domains

🛀
Surveillance

Abandoned Object Detector

Real-time detection of unattended luggage in public spaces (airports, stations). Uses YOLO + IoU tracker with center-history smoothing, audible/visual alerts, and optional Deep SORT integration.

YOLOv8 OpenCV ByteTrack NumPy RTSP
View Repository →
👨‍👩‍👦
Anomaly Detection

Crowd Movement Monitor

Anomaly detection in dense crowds via dense optical flow + MAD statistics. Detects panic, stampedes, and surges. Optional Qwen2-VL semantic verification with a live overlay dashboard.

YOLO Optical Flow Qwen2-VL NumPy OpenCV
View Repository →
🫝
Healthcare / Safety

Fall Detection System

Keypoint-based fall detection for elderly care and hospitals using YOLO pose estimation. Implements a temporal state machine (STANDING → FALLING → DOWN → ALERT) with webhook notifications and VLM second-opinion verification.

YOLO Pose PyTorch Webhooks Qwen2-VL RTSP
View Repository →
🔥
Surveillance / Safety

Fire Start Detection

Early-stage fire & smoke detection on CCTV feeds. Custom YOLOv8-nano model with consecutive-frame confirmation to suppress false positives. Optional Qwen2-VL verification and SMTP email alerts with cooldown.

YOLOv8-nano Qwen2-VL OpenCV SMTP Custom Model
View Repository →
👁
Biometrics

Face Recognition Surveillance

Real-time facial recognition pipeline using InsightFace (SCRFD detector + ArcFace 512-d embeddings) for automated face matching. Flask SSE web dashboard, webhook notifications, and GPU/CPU ONNX Runtime support.

InsightFace ArcFace ONNX Flask SSE OpenCV
View Repository →
🎙
Anomaly Detection

Video Anomaly Detection (STEAD)

Spatiotemporal anomaly scoring using X3D-S backbone + STEAD model with Performer attention. Processes 16-frame clips, applies moving-average smoothing, and triggers alerts via snapshot + webhook + optional Qwen2-VL explanation.

PyTorchVideo X3D-S STEAD Qwen2-VL Transformers
View Repository →
🚨
Public Safety

Weapon Detection System

Real-time firearm detection for CCTV deployments. YOLOv8 detection with temporal persistence logic + Qwen2-VL structured threat assessment (weapon type, holder description, threat level 0–100, rationale). 4-bit VLM quantization support.

YOLOv8 Qwen2-VL PyTorch bitsandbytes CLAHE
View Repository →
🔄
Traffic / Surveillance

Wrong-Way Detector

Detects wrong-direction movement for persons & vehicles using YOLO + ByteTrack and trajectory analysis. Interactive calibration GUI, ROI polygon masking, gate-line crossing, CSV event logging, and Docker support.

YOLOv8 ByteTrack Shapely Docker YAML Config
View Repository →
⚙
Prompt Engineering

AutoPromptOptimizer

Automated prompt optimization loop using local LLMs via Ollama. Iteratively reformulates user prompts, scores candidates using an LLM-as-judge framework, persists best prompts to JSON, and logs all iterations to CSV. Includes an A/B testing harness for comparing prompt variants.

📈 LLM-as-judge scoring
🔄 Iterative refinement loop
📊 CSV iteration logging
Python Ollama LLM-as-judge Prompt Engineering JSON / CSV
View Repository →
📚
RAG / LLM Application

GenAI RAG Assistant

Complete Retrieval-Augmented Generation system with FastAPI backend and Streamlit UI. Features hybrid retrieval (FAISS dense + BM25 lexical), RRF fusion, conversation memory, generative indexing with multi-query rewriting, and LLM-based evaluation endpoints. Fully local - no cloud dependencies.

🔍 Hybrid retrieval + RRF fusion
🤖 Local LLM via Ollama
💬 Conversation memory (3-turn)
FastAPI FAISS sentence-transformers Streamlit Ollama pypdf
View Repository →

Get In Touch

Open to collaborations, freelance projects, and opportunities in AI engineering.

GitHub @kyrrus
↗
LinkedIn Akram-Hanouz
↗
Email akram.hanouz@gmail.com
↗
Available for Work & collaboration