MLOps Developer at Syslab.ai (2025-08 – 2026-1)
EyeconAI — Production Multi-Camera AI Video Analytics Platform
- Conceived and launched EyeconAI, a real-time video-analytics platform for hospital use cases, transitioning from manual CCTV review to automated, evidence-backed safety alerting across 6+ camera feeds.
- Curated and labeled camera-specific datasets in Roboflow, then iterated across multiple YOLOv8/YOLO11 architectures and dataset versions to land production detectors — a 4-class exit-check model (child, guard, slip, attendee) reaching 93.6% mAP50, a 6-class ward-monitoring model (attendee, patient, staff, gloves/no-gloves) reaching 97.5%+ mAP50, and a tray/trolley detector used for event triggering.
- Migrated cost- and reliability-sensitive anomalies from Gemini to custom-trained YOLO models after hitting hallucinations, API rate-limits, and high CPU/RAM utilisation in production — cutting cost and improving reliability.
- Routed anomaly handling through a hybrid pipeline that uses Gemini for context-heavy calls (like people exiting with document-checks and guard presence) and local YOLO rules for simpler ones (like trolley stationarity) — with tray detection triggering Gemini only when a cannulation was actually starting, and filtering out false positives from bystanders, occlusion, and brief detection loss throughout.
- Assembled the event-driven backend spanning RTSP camera ingestion, Kafka streaming, per-camera Python inference workers, Express APIs, PostgreSQL, and a Next.js dashboard — enforcing evidence-required alert persistence so every flagged incident includes a reviewable clip, not just a model verdict.
- Containerized an eight-service Docker Compose stack behind Nginx/TLS with health checks, log rotation, and daily database backups; engineered resilient RTSP-over-TCP camera networking with a bounded HTTP fallback; and diagnosed/resolved live incidents including NVR firewall restrictions and a GCP static-IP network-tier mismatch.
Agentic AI-Based Hiring Platform
- Architected end-to-end ML pipelines for Uhire, an agentic AI recruitment platform — integrating multilingual speech-to-text (Whisper), LLM-based psychometric scoring, CV parsing, and Urdu transcription.
- Developed a document edge-detection stack — YOLOv8 bounding-box model to detect, deskew, and crop CNICs, educational certificates, CVs, and other applicant documents ahead of OCR and verification.
- Shipped production-grade face verification APIs using FaceNet, MTCNN, and InceptionResnetV1 with similarity-threshold optimization, achieving 90% 1:1 verification accuracy.
- Provisioned scalable, low-latency inference infrastructure on Google Cloud with HTTP(S) Load Balancer and autoscaling (99% uptime under heavy load), and reduced model deployment cycle time by 30% through standardized ML workflows and CI/CD automation.
Research Assistant at National University of Computer and Emerging Sciences (2024-08 – 2025-07)
Supervisor: Dr. Jawwad Ahmed Shamsi (PI) and Dr. Burhan Khan (Co-PI)
- Led first-author IEEE publication of the manuscript titled SAFE under an HEC-funded smart city surveillance project, developing a scalable and context-aware ML framework for mixed-critical event detection.
- Trained and optimized LSTM sequence models and a Vision Transformer for real-time fire and traffic incident classification (85% accuracy after tuning), then deployed the pipeline on Google Cloud with sub-1s inference latency and 80%+ accuracy across three critical event levels.
- Scaled the system to support 400+ concurrent users by implementing adaptive framerate monitoring and resource optimization, reducing cloud infrastructure costs by 70%.
