Aaron Patrick Nathaniel

Senior AI Engineer · AI Lead · Malaysia, remote

Aaron Patrick
Nathaniel

Nine years taking computer vision, LLM, and RAG systems from research to production, and now building recommendations at Mindvalley. I lead small teams and I measure everything.

In production AI
9+ years
Granted patents
3
Face model training
15M images
Face search
10K to 15K a day
Deployment time
2.5 h → 20 min
Team led
4 people

WHAT I
BRING.

  1. 01

    Production computer vision

    Face recognition, face search, document authentication, and on-device models, including three granted patents.

  2. 02

    LLMs, RAG, and fine-tuning

    Enterprise RAG used by about 100 employees, LoRA fine-tuning, and Decant, a study of what actually works.

  3. 03

    Evaluation you can trust

    Held-out exams, noise floors, two scorers on every answer, and the negative results kept in.

  4. 04

    Leading small teams

    Led two junior engineers and two interns at once. Code reviews, onboarding, interviews, and the quarterly roadmap.

SYSTEMS
IN PRODUCTION.

Client names stay private. The engineering does not.

Identity document authentication: Phone video or images, then Classify document, then Microprint, ghost image, hologram vs glare, then Combine the checks, then Genuine or reject Phone video orimages Classifydocument Microprint,ghost image,hologram vsglare Combine thechecks Genuine orreject
Problem
Onboarding has to tell a genuine identity card from a printed or edited copy, from a phone camera, in seconds.
What I built
The authentication methods behind three granted patents: microprint checked on contrast-enhanced images; a pipeline that classifies the document, compares the ghost image with the photo, checks colour capture, and finds the security hologram; and a hologram check from live video that uses where the hologram sits relative to the glare to reject printed copies.
Decisions
  • Combine several independent checks, so one forgery trick is never enough.
Stack
  • Python
  • OpenCV
  • Deep learning classifiers
My role
Co-inventor on all three. See the patents ↓

Face recognition model: 15M cleaned images, then Crop and align (new cropping), then Train, then Internal benchmark vs vendors, then NIST submission, then Production 15M cleanedimages Crop andalign (newcropping) Train Internalbenchmarkvs vendors NISTsubmission Production Iterate on data and crops
Problem
Off-the-shelf and paid face recognition was not accurate enough on the faces and capture conditions the product actually saw.
What I built
An in-house face recognition model. About a year of cleaning data and training on 15 million images, reaching about 98% 1:1 accuracy on our internal benchmark. On that benchmark, at the time, it beat InsightFace and the paid solutions we evaluated, with far less training data than the big vendors.
NIST 1:N
Tested in NIST's 1:N identification evaluation, alongside the largest face recognition vendors in the world. In mugshot searches across a 1.6 million face gallery, the right person came back as the top match 91% of the time.
What came next
My research direction, a new face cropping approach and more mugshot-style training data, drove the gains in the company's September 2026 NIST submission: roughly 30 to 40% fewer missed matches on border and visa photos.
Decisions
  • Invest in data cleaning before reaching for new architectures.
  • Own the model instead of paying per call for a vendor API.
Stack
  • Python
  • PyTorch
  • OpenCV
My role
Led the research, from data cleaning to training.

Face search at scale: Image in, then Detect and align, then Face embedding, then FAISS index, then Top-k candidates, then Threshold, then Review or alert Daily new images Image in Detect and align Face embedding FAISS index Top-k candidates Threshold Review or alert
Problem
Fraud teams need to know whether a new applicant's face has been seen before, across a growing archive, every day.
What I built
A production face search service on FAISS, built on the in-house recognition model for its embeddings.
Decisions
  • A FAISS vector index, so a growing archive stays searchable every day.
  • Thresholds tuned against the review workload.
Stack
  • Python
  • PyTorch
  • OpenCV
  • FAISS
  • FastAPI
  • Docker
My role
Built the service.

Vision on the edge: Camera, then On-device model under 100 MB, then Liveness or detection, then Result OFFLINE, NO NETWORK Camera On-device model under100 MB Liveness or detection Result
Problem
Mobile and browser checks had to run fully on the device, with no connection.
What I built
Production vision models compressed to under 100 MB each for fully offline mobile inference, face monitoring tools for fraud detection, and object detection in the browser with WebAssembly.
Decisions
  • Compress the models to fit on a phone instead of calling a server.
  • Run detection in the browser through WebAssembly, so nothing has to be uploaded.
Stack
  • Python
  • OpenCV
  • WebAssembly
  • On-device inference
My role
Built the face monitoring tools and the WebAssembly detection.

Document intelligence: Card image, then Classify document, then Detect and read text (CTPN + CRNN, then PPOCR), then NER, then Structured fields Card image Classifydocument Detect and readtext (CTPN +CRNN, thenPPOCR) NER Structuredfields
Problem
Pulling personal details off a Malaysian identity card (MyKad) had to be fast enough for live onboarding.
What I built
Replaced a CTPN text detector and CRNN recogniser pipeline with PPOCR, cutting inference from about 4 seconds to about 1.5 per card. Also image classification for document types, and bank statement processing.
Decisions
  • Add NER after OCR, so the output is structured fields instead of raw text.
Stack
  • PPOCR
  • CRNN
  • CTPN
  • NER
  • Python
My role
Rebuilt the text extraction pipeline.

Enterprise LLM and RAG: Documents, then Chunk, then Embed, then Vector store, then Retrieve, then LLM, local or API, then Guardrails, then Answer Documents Chunk Embed Vector store Retrieve LLM, local or API Guardrails Answer Evaluate, then tune chunking and retrieval
Problem
Employees needed reliable answers from the company handbook and documents.
What I built
Internal LLM and RAG applications on open-source models: a handbook assistant, and a bank statement and document classification application. LoRA fine-tuning on the handbook, plus chunking, retrieval, and parameter tuning. About 90% accuracy and about 60% fewer hallucination and guardrail issues, and the custom LoRA models matched leading provider performance for the business use case.
Decisions
  • Measured local LLM deployment against the GPT API on cost and performance before choosing.
  • Hallucination testing and guardrails built in, not added afterwards.
Stack
  • Python
  • Ollama
  • LoRA
  • FAISS
  • Streamlit
My role
Led development of the production LLM applications.
Next
Decant, below, is the tool I wished I had while building this.

MLOps platform: Data prep change, then DVC and GitHub commit, then Jenkins, then Train, then MLflow registry, then Docker, then FastAPI, then Deploy Data prep change DVC and GitHub commit Jenkins Train MLflow registry Docker FastAPI Deploy
Problem
Training and deployment were manual and slow, and they depended on one person's machine.
What I built
DVC for data versions, GitHub for code, Jenkins for automation, and MLflow for experiments and the model registry. Docker optimised and services moved to FastAPI for faster inference.
Decisions
  • Every model traces back to the exact data and code that made it, so anyone can retrain or roll back.
Stack
  • DVC
  • GitHub
  • Jenkins
  • MLflow
  • Docker
  • FastAPI
My role
Introduced the company's first full MLOps pipeline.

Recommendations: Catalogue and user signals, then Candidate generation, then Ranking, then Content mix, then Safeguards, then Feed Catalogueand usersignals Candidategeneration Ranking Content mix Safeguards Feed Evaluation metrics feed the next ranking
Problem
Help people find the right next thing across several digital products, and keep doing it well over the long term.
What I built
Personalised recommendation systems: candidate generation, ranking, content mix, evaluation metrics, and safeguards for relevance, discovery, and long-term user outcomes. In progress, so no public metrics yet.
Decisions
  • Evaluation metrics and safeguards are part of the pipeline from the start.
My role
Senior AI Engineer on recommendations.

3 GRANTED
PATENTS.

  • US12026932B2

    Method to determine authenticity of security hologram

    Filed 2021. Granted 2 July 2024. United States.

    Co-inventor

    VIEW ON GOOGLE PATENTS ↗
  • MY192715A

    Method for authenticating identification documents

    Filed 2020. Granted 2022. Malaysia.

    First-named inventor

    VIEW ON GOOGLE PATENTS ↗
  • MY184165A

    A method to verify authenticity of a Malaysian identification document

    Filed 2019. Granted 2021. Malaysia.

    Co-inventor

    VIEW ON GOOGLE PATENTS ↗

DECANT.
PROVE IT WORKS.

I built Decant to turn a small open model into a specialist on your own documents: synthetic training data, LoRA fine-tuning, retrieval, tools, guardrails, and a held-out final exam. Then I tried to prove it worked.

Manuals
8
Student model
Qwen3-4B
Exam questions
285
Experiments
26+
Decant guardrails step: topics the model must decline.

Round 2 · 285 questions · independent judge

Retrieval did the work. Fine-tuning did not.

Untuned, no retrieval 18.8%
Fine-tuned + retrieval 83.8%
Untuned + retrieval 91.2%
Untuned + cross-encoder reranker 93.8%
Factual accuracy on 8 plant-safety manuals. Sampling noise alone moves a score by up to 6.2 points, so smaller gaps are never ranked.

Round 2 · general ability

Fine-tuning made it forget. One mix fixed it.

Base model 88.2%
After fine-tuning 35.3%
Fine-tuning + 20% general data 91.2%
Mixing in about 20% general instruction data restored base-model ability. The forgetting came from narrow training data, not from the model being small.

Negative results I kept

Two things that did not work.

  • 95.5 → 91.0

    Hybrid retrieval hurt

    Dense plus BM25 scored lower than dense alone.

    Round 1
  • 85.0% vs 78.7%

    Three examples beat the adapter

    Few-shot prompting out-scored the fine-tuned model.

    Round 1

How it was measured

Honest numbers, or none.

  • Report the noise floor, and never rank gaps smaller than it
  • Two scorers on every answer, shown side by side
  • Measure under the budget you ship with

Retrieval is how the model knows things. Fine-tuning is how it writes.

READ THE
FULL STUDY ↗

HOW I
LEAD.

I led a team of four at Innov8tif: two junior engineers and two interns. Fast models were only half the job. The other half was making sure the team could run, fix, and retrain them without me in the room.

  1. 01

    One pipeline

    OptimisedChange the data prep, commit, and training runs on its own to a tracked, deployable model.

    MaintainableI taught the whole team the workflow, so training no longer depended on one person's machine.

  2. 02

    Code reviews

    OptimisedEvery change reviewed for speed: smaller models, faster inference.

    MaintainableAnd for whether the next person can read, test, and change it.

  3. 03

    Onboarding and hiring

    OptimisedPrepared the technical interviews, to hire for the work.

    MaintainableNew engineers and interns learned the same pipeline and standards as everyone else, so nothing depended on one person.

  4. 04

    Quarterly roadmap

    OptimisedPlanned each quarter with the team.

    MaintainablePeople built what they had helped plan, and owned it after release.

The test of a good system is the second person who has to maintain it.

FROM 2018
TO NOW.

  1. 01

    2015 to 2018

    Multimedia University

    Bachelor in Computer Science, Data Science.

    3.74CGPA

  2. 02

    2018

    Innov8tif, Full-Stack Engineer

    Computer vision for fraud detection and eKYC, in Python and OpenCV.

    2018Joined Innov8tif

  3. 03

    2019

    First patent filed

    Microprint verification for Malaysian identification documents. Granted in 2021 (MY184165A).

    Patent 1Malaysia

  4. 04

    2020

    Document authentication patent

    Authenticating identification documents with several algorithms and deep learning. Granted in 2022 (MY192715A), as first-named inventor.

    Patent 2Malaysia

  5. 05

    2021

    US patent filed

    Judging whether a security hologram is real from live video. Granted in the United States in 2024 (US12026932B2).

    US patentGranted 2024

  6. 06

    2026 to now

    Mindvalley, Senior AI Engineer

    Personalised recommendations across multiple digital products.

    NowRecommendations

SKILLS.
END TO END.

Languages and ML

  • Python
  • TypeScript
  • JavaScript
  • C++
  • SQL
  • PyTorch
  • TensorFlow
  • Hugging Face
  • OpenCV
  • YOLO
  • PPOCR
  • NER
  • Scikit-learn

LLM and GenAI

  • OpenAI
  • Gemini
  • Ollama
  • LangChain
  • RAG
  • LoRA fine-tuning
  • FAISS
  • Guardrails
  • Prompt evaluation

Agentic AI

  • MCP
  • Tool and function calling
  • Multi-agent orchestration
  • Agent workflows
  • Codex
  • Claude Code
  • OpenCode

Product and infra

  • FastAPI
  • React
  • Next.js
  • Docker
  • Kubernetes
  • Airflow
  • MLflow
  • DVC
  • GitHub Actions
  • Jenkins
  • GitLab CI/CD
  • AWS
  • GCP
  • Azure
  • Playwright
  • Vitest
  • PyTest
  • Stripe

What I can own end to end: requirements and scoping, data prep and knowledge-base design, model and provider evaluation, RAG design, retrieval tuning and fine-tuning, hallucination testing and guardrails, backend APIs, auth and frontend, testing, Docker, CI/CD and MLOps, then monitoring, feedback loops, and production support.

ALSO
BUILDS.

I ship products on my own time. They are proof of speed and range, and each one has a full section on the portfolio.

Let's talk.

Open to remote Senior AI Engineer and AI lead roles.