Nouman Ahsan

Machine Learning Engineer

Recognition • Full-time

Jan 2023 - Present • 3 yrs 9 mos

Design and ship production AI systems end-to-end across computer vision, OCR, document intelligence, generative AI and LLMs, from model training and optimization to scalable backend services and on-device SDKs. - Face Recognition Platform: event-driven biometric system with Kafka-buffered inference and HNSW vector search in pgvector; 2,000 operations with zero failures, p95 latency around 5.7 s. - Watchtower: distributed video analytics platform where a central controller orchestrates a fleet of edge workers for real-time detection and tracking, with automated failure recovery. - LLM Document Understanding: typed field extraction with per-field confidence scores across 14 document types, plus grounded chat that answers questions from the document itself. - ID Document Reader: reads GCC national IDs, passports from around 50 countries and UAE visas, with a server-side ONNX pipeline and a fully offline Android and iOS SDK. - Live Face Capture SDK: browser-native capture with randomized liveness challenges, designed so no biometric frames ever leave the user's device. - Custom OCR Models: lightweight CRNN models for English and Arabic with 98%+ precision and recall, small enough for edge deployment. - Signature Extraction: pipeline to extract and enhance signatures from ID cards, reaching around 0.92 SSIM. - Real-time monitoring tools and dashboards giving operations teams live visibility into high-volume processing workflows. - Offline LLM tool for querying and documenting large codebases. Work spans the full lifecycle: requirements, architecture, model development, API design, deployment with Docker, and production monitoring.

Associate Machine Learning Engineer

iENGINEERING • Full-time

Nov 2020 - Oct 2022 • 1 yr 11 mos

Worked with Transconomy, the Applied Research department of iENGINEERING, building FHWA-standard roadway asset inventories - an end-to-end production pipeline from field data capture through automated model inference to QC'd deliverables. • Architected and trained deep learning models - object detection, instance and semantic segmentation, classification - for traffic signs, lane markings, and pavement distresses. • Built automated data pipelines covering the full inventory lifecycle: ingestion, model inference, and delivery of roadway geodatabases (ArcGIS Pro, QGIS) with automated MS SQL processing. • Coordinated a 35-person data annotation and QC team - forming project groups, assigning and distributing work, and reviewing QC output. • Deployed models across AWS cloud and NVIDIA Jetson edge devices. • Conducted technical interviews and mentored new engineers.