AU

日本語での自己紹介

はじめまして。アラン・エマニュエル・ウマリと申します。フィリピン・マニラを拠点に、応用AIとコンピュータビジョンの研究・開発に取り組んでおります。現在はKDCI(Key Discovery Consulting, Inc.)でAIチームを率い、LLMエージェントや音声AIの実用化を担当しております。

研究では、単眼の交通カメラ映像から交差点の方向別交通量を計測し、その結果を自ら検証する手法「GATE」を開発しました。本研究は、神戸で開催されるIEEE FMLDS 2026において、2026年11月21日に発表いたします。

今後は日本の大学院で研究を続けたいと考えております。研究室でのご指導や受け入れについて、ご相談させていただけましたら幸いです。

ご連絡はこちら

About

I’m an AI specialist and the founding engineer of the AI team at KDCI (Key Discovery Consulting, Inc.). I take AI systems from first idea to production: voice and chat agents, LLM agents with evaluation and human review built in, and data pipelines that classify and check records at scale. I also train staff to use AI well.

Before KDCI, I built edge computer-vision systems as an AI engineer at Supervaise, and spent six years at De La Salle University’s Intelligent Systems Laboratory, from research assistant to project lead. I’m a licensed electronics engineer (ECE, ECT), partway through an M.Sc. in Data Science at DLSU, and I hope to continue my graduate studies in Japan.

Research

  1. GATE: Graph-Admissible Transition Evaluation for Self-Auditing Turning-Movement Counting from Monocular Traffic Cameras

    A. E. B. Umali et al. 2026 IEEE International Conference on Future Machine Learning and Data Science (FMLDS), Kobe, Japan. First author and presenter.

    Accepted, to appear in the proceedings

  2. Smart Shelf System for Customer Behavior Tracking in Supermarkets

    J. A. C. Jose, C. J. B. Bertumen, M. T. C. Roque, A. E. B. Umali, et al. Sensors, 24(2), 367, 2024. Built from my undergraduate thesis (Gold Thesis Nominee): I worked on the weight-sensor system, the off-the-shelf vision system and matching customers to products.

Read my academic CV

Where my work meets research

My day job is applied AI. Each project raises a question I’d like to study properly, ideally with a lab in Japan.

Grounded RAG and voice agents

How do you keep answers grounded when the stakes are real?

Retrieval-augmented, tool-using agents that answer only from approved sources, and hand over to a person when they can’t.

Grounded in Elaine: every price from a tool

Discuss this together

Selected work

Applied AI at KDCI, and computer vision before it. Client projects are described without naming the client.

Elaine, KDCI’s voice agent

Answers KDCI’s phone line and website, captures leads, and hands callers to a person when needed. Live about four weeks after the request. Every quoted price comes from a pricing tool, never from the model.

  1. Caller
  2. AI agent
  3. Pricing tool
  4. Person or sales

Research angle Keeping voice RAG grounded

ElevenLabs, Twilio, Make

CoreIndex, KDCI’s AI-vendor directory

The data platform behind KDCI’s AI-vendor directory on kdci.ai. It aggregates about 34,000 providers and 100,000+ public reviews, and 1,500+ providers are published. A 250-sample model A/B test raised verified evidence from 86.9% to 99.0% at about 5× lower cost.

  1. Web data
  2. Classify
  3. Evidence check
  4. Publish

Research angle Small test sets that pick the right model

Python, LLMs, Batch API, automated tests

Support-reply agent for an online retailer

Drafts replies to customer tickets for a US e-commerce client. Graded accuracy rose from 36% in the first graded week to 97%, holding above 90% monthly. A person approves every reply, and anything with financial risk is escalated.

  1. Ticket
  2. AI draft
  3. Risk check
  4. Human approval

Research angle Human-in-the-loop evaluation

LLM agent, guardrails, human review

AI Foundations training and responsible-use rules

Co-led KDCI’s AI Foundations pilot as its top-rated trainer (4.98/5, NPS +90), and authored the company’s AI data-governance framework: what is safe to paste into AI tools, what to redact first, and what never goes in.

  1. Learn
  2. Practice
  3. Use safely

Research angle How people learn to work with AI

Training, AI governance

License-plate recognition at the edge, Supervaise

Re-architected a plate-recognition pipeline from 108.6 s to 11.2 s runtime and stress-tested it at up to 445 plates a minute on NVIDIA Jetson devices. Automated retraining replaced a 5–8 hour manual workflow with one command, and foundation-model pre-annotation halved labelling time.

  1. Camera
  2. Detect
  3. Read plate
  4. Store

Research angle Efficient, reliable vision on low-power devices

NVIDIA Jetson, TAO, YOLO, Kafka, AWS

Experience

  1. Jan 2026 – now

    AI Specialist (AI Team Lead)

    KDCI (Key Discovery Consulting, Inc.)

    Founding engineer of the AI team. Takes agentic-AI systems to production with evaluation and guardrails, and onboards the growing team.

  2. Jan 2025 – Jan 2026

    Artificial Intelligence Engineer

    Supervaise, Inc.

    Edge computer vision and MLOps for license-plate recognition on NVIDIA Jetson.

  3. 2020 – Jan 2026

    Project Technical Associate, research and teaching assistant

    De La Salle University, Intelligent Systems Laboratory

    Led an AI vehicle-counting and traffic-analytics project, co-supervised research teams totaling 8 members, and helped teach 500+ engineering students.

  4. 2022 – Jan 2026

    Founding team member and technical lead

    NVIDIA AI Philippines

    A community of 2,000+ members. Co-taught NVIDIA Deep Learning Institute workshops, including one for 252 participants from 17+ universities.

  5. In progress

    M.Sc. Data Science

    De La Salle University

    Coursework partway complete. Planning to continue graduate study in Japan.

  6. 2023 · 2024

    B.Sc. Electronics and Communications Engineering · Licensed ECE and ECT

    De La Salle University

    Gold Thesis Nominee; the thesis became the Sensors paper. Board exam 85.70%, among the top ratings of DLSU takers. National Champion, IECEP Interscholastic Quiz Show 2022.

See my one-page resume

Tools I work with

  • Python
  • PyTorch
  • OpenCV
  • NVIDIA Jetson
  • Claude
  • ChatGPT
  • Gemini
  • ElevenLabs
  • Twilio
  • Make
  • Hugging Face
  • GitHub

Resume and CV

Two ways in, depending on why you’re here.

For professors and scholarship committees Academic CV 2 pages 1st authorIEEE FMLDS 2026, Kobe Sensorsjournal paper, 2024 85.70%ECE board exam, among the top ratings of DLSU takers

Read the academic CV on its own page, or download the PDF.

For teams and collaborators Resume 1 page 36% → 97%graded accuracy of an AI support agent ~4 weeksfrom request to a live voice agent 108.6 s → 11.2 splate-recognition pipeline runtime

Read the resume on its own page, or download the PDF.

Need the complete record, with every role, award and certification? Request my full CV

Collaborate with me

Tell me what you have in mind: a research collaboration, graduate study or supervision, a consulting or proof-of-concept project, or a talk. I reply personally, usually within a few days.

Ask my AI

Soon you’ll be able to ask an AI assistant about my research and experience, and how I can help with your work, by chat or voice. Until then, send a collaboration request and I’ll answer myself.