Tech Lead, Applied AI, Tagteam Technologies
Jan, 2024 - Present
Architected and engineered a new retail analytics product that processed over 9,000 hours of CCTV footage, deploying a distributed video processing pipeline on GKE utilizing Temporal and Kafka to natively manage complex multi-model asynchronous dependencies and auto-scaling container configurations.
Extended the platform to support offline batch video ingestion for a Singapore Tourism Board deployment, engineering a manual upload pipeline and developing an OCR-based parsing module to extract frame-embedded timestamps for automated metadata start-time synchronization.
Designed the product's core computer vision pipeline by adapting SAM2 for high-throughput online multi-object tracking outperforming ByteTrack, fusing person detection, custom-finetuned behavioral models, and demographic estimation into a unified real-time metadata generation service.
Developed the platform's multi-tenanted architecture, building a fully customizable Vue 3 web dashboard with dynamic Apache ECharts configurations to securely serve diverse clients; integrated an embedded conversational agent to assist users in interpreting complex data visualizations directly within the UI.
Led the technical execution of a safety-critical runway anomaly detection system for an international airport, optimizing ONNX/TensorRT inference on Jetson AGX Orin under severe vehicle power constraints to achieve real-time hazard detection; integrated an event-driven MQTT framework with remote API revocation and a live operations dashboard for real-time visualization.
Engineered an autonomous multi-agent backend using Python, Vertex AI, and Google Agent Dev Kit, implementing Langfuse for LLM observability; utilized Holt-Winters forecasting to establish predictive intervals, flagging out-of-bound deviations to automate root-cause analysis and generate actionable insight reports for teams lacking dedicated business analysts.
Engineered and delivered an image search system proof-of-concept for a marketing agency, integrating a lightweight image captioning model, hybrid search algorithms, and cross-encoder reranking to achieve high-accuracy natural language retrieval of past collaterals using Weaviate.