I lead the AI team at AI Lens Group in Kuala Lumpur. We build software that reads bank statements, payslips and other financial paperwork, so lenders can assess borrowers properly, including small businesses with little credit history. Before that I spent five years on a Ph.D. teaching models to read text in the dark.
16 people on the AI and engineering team I lead at AI Lens Group
50,000+ financial documents a month read by our platforms, for banks and lenders in Malaysia
EMNLP 2026 for PRISM, my first-author paper, which also won a Poster Paper Award at FinLLM @ IJCAI 2026
A map of this site
Each dot is a page or a project. Click or tap one to open it.
I lead a team of 16 that builds the software banks and lenders in Malaysia use to read financial documents and score credit. It reads bank statements, payslips, identity cards, invoices, financial statements and credit bureau reports, more than 50,000 a month, at over 95% accuracy and under five seconds a page, on the lenders' own servers. The business brings in seven-figure annual recurring revenue (MYR).
I also run our research. Our 3rd place in the ACM ICAIF 2025 retrieval challenge grew into PRISM, my first-author paper at EMNLP 2026, and what we learn there feeds back into what we ship.
More
The stack is Python and FastAPI, React, PostgreSQL, RabbitMQ and Docker, running on Azure and on-premise. I set up our Azure infrastructure and the GitHub Actions pipeline (self-hosted runners, branch gating, release checks), plan the sprints, and mentor engineers across backend, DevOps, OCR and AI system design.
Vice President of Engineering, October 2023 to March 2024
Etch Media runs advertising on LED display boards. Cameras recognized the cars going past (make, model, color) and the people nearby (age, gender, emotion), and an LLM matched what they saw to ad categories, so the board could play a fitting ad in real time.
I built the C# player that schedules and switches content on the custom LED boards, the backend services that connect the AI models to ad delivery, and the CI/CD on Azure.
I taught hands-on OCR classes (video) on reading license plates and shop signboards with PaddleOCR and MMOCR. The course materials are open source: license plates and signboards.
2023
AkiraKan
Part-time ML/AI Research Scientist, January 2023 to September 2023
I worked on OCR for traffic-scene photos, reading lamp post IDs and license plates, and on some 3D projects.
Part-time Research Intern, March 2020 to December 2022
I worked on reading the characters printed on microchips, which are often low-contrast, blurred or broken. With Baidu and partner universities we built ICText, the largest dataset of text on integrated circuits, and ran it as an ICDAR 2021 competition.
More
A curriculum-learning training method I proposed for ICText improved character recognition and led to two papers. I also helped undergraduate interns with image classification and object detection projects.
Research Assistant, September 2018 to December 2023
My Ph.D. years, in Prof. Chan Chee Seng's lab. I worked on finding and reading text in photos (detection, recognition and spotting) and on brightening low-light images, and helped build the ArT, EST-VQA and ICText benchmarks.
Intern Software Engineer, February 2017 to August 2017
My first job: fixing bugs, building customer requests and writing tests for a bank loan system, in the Centre of Excellence team.
Research
I started out in computer vision, teaching models to read text in awkward places: curved shop signs, tiny print on microchips, photos taken almost in the dark. These days the awkward place is financial paperwork. A few projects I'm proud of, newest first:
PRISM: finding the right page in a financial filing
2025 to 2026, AI Lens Group. EMNLP 2026 Industry Track
Analysts ask questions like "what was the operating margin last year?", and the answer sits on one page of a filing hundreds of pages long. Plain vector search often misses it. PRISM gets there without any training: a tuned system prompt, worked examples chosen for each question, and a retrieve, rerank and refine step before a single LLM ranks the results. We also scored every run on accuracy, latency and cost together, because that trade-off is what decides whether something ships. It started as our 3rd-place entry in the ACM ICAIF 2025 retrieval challenge and later won the Poster Paper Award at the FinLLM Symposium @ IJCAI 2026.
2025 to 2026, AI Lens Group. ICLR 2026 workshop; earlier version an oral at ACM ICAIF 2025
Many Malaysian small businesses are doing fine but have too little credit history for a bank to score. Their bank statements tell the real story. AI-BAAM reads six months of statements, pulls out the account details and every transaction, checks the PDFs for tampering, maps who the money moves between, and turns the cash flow into features a credit model can use. Loan officers stay in the loop at every step.
Ph.D. work, University of Malaya. ICPR 2022; Signal Processing: Image Communication 2025
My Ph.D. asked a simple question: does OCR still work on photos taken almost in the dark? Mostly it doesn't. Brightening the image makes it look nicer but leaves the text a smudge. So we built Text in the Dark, the first extremely low-light text dataset, with 3,075 images and 59,975 text instances annotated, and an enhancement method that restores the text itself, not just the brightness.
Annotated scenes from the Text in the Dark dataset.
2020 to 2024, ViTrox and University of Malaya. ICDAR 2021 competition; Pattern Recognition
The markings on a chip are tiny, low-contrast and often scratched or broken, and factories need to read and inspect every one. With ViTrox, Baidu and partner universities we built ICText, the largest dataset for it, with 10,000 images and 100,152 characters annotated, and ran it as a competition at ICDAR 2021.
A chip from ICText, every character boxed and labelled.
EMNLP 2026, Industry Track. PRISM: Prompt-Refined In-Context System Modeling for Financial Retrieval.
arXivcode
Poster Paper Award, FinLLM Symposium @ IJCAI 2026.
Oral, 9th Workshop on Machine Learning in Finance @ KDD 2026.
ICIP 2023. Rethinking Long-Tailed Visual Recognition with Dynamic Probability Smoothing and Frequency Weighted Focusing.
papercode
ICPR 2022. Extremely Low-Light Image Enhancement with Scene Text Restoration.
papercode
ICDAR 2021. Competition on Integrated Circuit Text Spotting and Aesthetic Assessment.
paperdataset
CVPR 2020. On the General Value of Evidence, and Bilingual Scene-Text Visual Question Answering.
paperdataset
ICDAR 2019. Robust Reading Challenge on Arbitrary-Shaped Text (RRC-ArT).
paperdataset
ICDAR 2019. Competition on Large-Scale Street View Text with Partial Labeling (RRC-LSVT).
paperdataset
ACM MM 2018. Unprecedented Usage of Pre-trained CNNs on Beauty Product.
paper
All journal papers
Signal Processing: Image Communication, 2025. Text In the Dark: Extremely Low-Light Text Image Enhancement.
papercode
Pattern Recognition, 2024. When IC Meets Text: Towards a Rich Annotated Integrated Circuit Text Dataset.
papercode
Workshop papers
ICLR 2026 workshop. AI-BAAM: AI-Driven Bank Statement Analytics as Alternative Data for Malaysian MSME Credit Scoring.
arXiv
The 2nd Workshop on Advances in Financial AI @ ICLR 2026.
Earlier version: Cash Flow Underwriting with Bank Transaction Data: Advancing MSME Financial Inclusion in Malaysia (arXiv). Oral, AI for Financial Inclusion, Risk Modeling & Resilience in Emerging Markets @ ACM ICAIF 2025.
Collaborators
People I've worked closely with, and learned a lot from:
