Hanoi, Vietnam · UTC+7

Henry Le

Senior AI Engineer

Seven years shipping AI to production — computer vision, document intelligence, and agentic LLM systems.

I build AI systems that survive contact with real users: multimodal document pipelines, retrieval stacks running on-premise, and long-lived conversational agents with memory. Currently at Torilab, previously Samsung SDS R&D Center, VinBrain, and VCCorp.

Resume GitHub LinkedIn Email
Henry Le
Experience

Where I've shipped

  1. Apr 2025 — Present
    Current

    Senior AI Engineer · Torilab

    Hanoi, Vietnam

    • Architected and shipped a multi-persona conversational AI platform on LangChain, LangGraph, and the OpenAI Agents SDK — 10+ production avatars across companion and lifestyle-coaching products, serving 2K+ monthly active users.
    • Designed a multi-layer long-term memory system for persona agents, combining a persistent user profile, episodic semantic recall, and rolling session summaries. Enabled cross-session personalization and measurably increased average session length.
    • LangGraph
    • LangChain
    • OpenAI Agents SDK
    • Vector search
    • Postgres
  2. May 2022 — Apr 2025

    Senior AI Engineer · Samsung SDS R&D Center

    Hanoi, Vietnam

    • Led end-to-end design of a legal document intelligence platform, architecting a multimodal pipeline over PDFs and images using vision-language models. Deployed a secure on-premise RAG stack with a self-hosted LLM, reranker, and vector search — improving retrieval relevance and cutting processing latency.
    • Built an Intelligent Document Processing platform extracting key information and tables from diverse documents, using Key-Value Understanding, Table Recognition, and Document Reconstruction models on a Django, Celery, and RabbitMQ backend. Applied to digitalize data for several large companies, including a top-3 bank in Vietnam.
    • Developed an end-to-end MLOps pipeline with Kubeflow on Kubernetes, reducing model deployment time by 50% and enabling automated training for 10+ production models.
    • Built a Sales Boosting Tool using Key-Value Understanding and TensorRT inference, supporting 50K monthly transactions at 98% accuracy and under 1s per request — an 80% reduction in processing time for finance operations.
    • VLMs
    • RAG
    • TensorRT
    • Kubeflow
    • Kubernetes
    • Django
    • Celery
    • RabbitMQ
  3. Nov 2020 — May 2022

    Applied Scientist · VinBrain

    Hanoi, Vietnam

    • Built a system for reviewing medical device registration applications for the Ministry of Health, using document alignment, text detection, text recognition, and question-answering models. Optimized deployment via knowledge distillation, reaching 95% accuracy and cutting processing time by 80% versus manual review.
    • Developed a facial stroke alert system using face landmark detection and a classification model, reaching an F1 score of 90% to enable early intervention.
    • Built the eKYC system for the AIViCare app at 95% accuracy, reducing user verification time to under 5 minutes.
    • OCR
    • Document QA
    • Knowledge distillation
    • PyTorch
    • eKYC
  4. Jul 2019 — Nov 2020

    Computer Vision Engineer · VCCorp

    Hanoi, Vietnam

    • Built a 3D-CNN video recognition model detecting violent actions at over 90% accuracy, used to filter violent content on the Lotus social network.
    • Developed and maintained a MobileNetV3 model for sensitive-content image classification, sustaining 98% accuracy across user-uploaded content.
    • 3D-CNN
    • MobileNetV3
    • TensorFlow
    • Content moderation