Indra Imanuel

AI Research Engineer at Laputa, Seoul · Ph.D. in Computer Engineering

Hi! I'm Indra, an AI researcher and engineer based in Seoul. I work on vision–language models, currently on diagnosing and repairing fine-grained visual-discrimination failures in Mixture-of-Experts VLMs; and on multimodal document understanding for retrieval systems.

At Laputa I build Unibook's LLM-based multimodal Q&A system over Korean university textbooks, and I lead a research project on document-hierarchy repair that grew out of that work. Before this I spent over a year at STANS building real-time computer-vision models for intelligent CCTV, some of which were deployed at Seoul subway stations, and before that I did my M.Sc. and Ph.D. at Dongseo University with Prof. Suk-Ho Lee on generative models for image restoration.

I'm from Surabaya, Indonesia, and have lived in Korea since 2019. I'm currently looking for a postdoctoral position.

Research interests: vision–language models · Mixture-of-Experts and efficient architectures · fine-grained visual understanding · model diagnosis, repair, and evaluation methodology · multimodal document understanding · generative models for image restoration.

News

Sep 2026Preparing two first-author manuscripts: on answer-position bias in MoE VLMs (targeting ICML 2027) and on second-pass document-hierarchy repair (targeting ICDAR 2027).
Jan 2026Started an independent research project on expert routing and fine-grained visual discrimination in MoE vision–language models, with two former labmates.
Dec 2025Joined Laputa, Inc. (AI subsidiary of educational book publisher Sapyoung) in Seoul as an AI Research Engineer, building an LLM-based multimodal Q&A system over Korean university textbooks.
Nov 2025Gave a guest lecture on vision–language models — trends, applications, and implementation — for Indonesian undergraduates (ProjekinAja).
May 2025Spoke at the International Symposium on Management 2025 (University of Surabaya × Ho Chi Minh University of Banking) on AI applications and data generation.
Oct 2024Joined STANS, Inc. as an AI Research Engineer, working on intelligent-CCTV models: object detection, crowd counting, and violence detection.
Sep 2024Guest lecture for graduate students at the University of Surabaya on computer vision and image generation.
Aug 2024Received my Ph.D. in Computer Engineering from Dongseo University. Thesis: Image Compression and Demosaicing based on Denoising Diffusion Null-Space Model.
Aug 2024Paper on optimized color filter arrays for diffusion-based demosaicing published in IEEE Access.
Dec 2023Paper on demosaicing-based image compression with a channel-wise decoder published in the International Journal of Internet, Broadcasting and Communication.
Oct 2022Best Paper Award (Silver Prize) at IEEE ICCE-Asia 2022 for Image Compression with Channel-wise Decoder.

Selected research All publications

The Calibration Law: Diagnosing and Repairing Answer-Position Bias in Mixture-of-Experts Vision-Language Models

Indra Imanuel, et al.

Manuscript in preparation (ICML 2027 target) · independent project, Jan 2026 – present

MoE vision–language models fail badly at telling apart images that differ in one small detail. We show much of the underlying failure is a severe answer-position bias, introduce a zero-cost diagnostic for it, and find, across eight open MoE VLMs, that training-free interventions succeed in proportion to how well they calibrate that bias. A label-free recipe substantially improves the VisMin group score and transfers zero-shot to Winoground and ColorSwap.

Second-Pass Document Hierarchy Repair for PDF Extraction (working title)

Indra Imanuel, et al.

Manuscript in preparation (ICDAR 2027 target) · Laputa, 2026

Can a second pass repair the document hierarchy that PDF extractors get wrong? An end-to-end evaluation on three public benchmarks and Korean textbooks, a self-supervised pointer model trained without human labels, and LLM/VLM baselines.

Optimized Color Filter Array for Denoising Diffusion Null-Space Model-Based Demosaicing

Indra Imanuel, Hyoseon Yang, Suk-Ho Lee

IEEE Access, 2024 SCIE · Ph.D. thesis work

A color-filter-array pattern and optimization scheme designed jointly with a diffusion null-space demosaicing model, giving better demosaiced images and compression rates than fixed patterns.

Super-resolution with adversarial loss on the feature maps of the generated high-resolution image

Indra Imanuel, Suk-Ho Lee

Electronics Letters, 2022 SCIE · M.Sc. thesis work

A GAN training scheme that applies the adversarial loss on the feature maps of the generated high-resolution image, via pre-trained networks, instead of on pixels alone. It improved on prior GAN-based super-resolution and outperformed the state-of-the-art CVPR method on a real-world low-resolution benchmark at the time of publication.

Selected projects All projects

Unibook: multimodal textbook Q&A system

Laputa · 2025 – present

LLM-based RAG and agentic workflow around a 20B+ open-weight multimodal MoE model over Korean university textbooks: designed the agentic workflow, fixed and tuned the retrieval pipeline, redesigned the database, chose and validated the production model by measurement, and own the cloud-GPU deployment stack.

LLMRAGAgentic workflowvLLMQdrantDocker

AWAS-Insight: crowd counting and violence detection for intelligent CCTV

STANS · 2024 – 2025

AI models for AWAS-Insight, STANS's intelligent-CCTV platform. Crowd-density estimation fine-tuned on company data with auto-labeling pipelines and deployed at Seoul subway stations; violence detection with action-recognition and optical-flow models integrated with a real-time object detector, quantized for real-time inference.

Crowd countingAction recognitionTensorRTONNX

Construction-site real-time translation

STANS · 2025 · early-stage R&D

Translating Korean construction jargon across multiple Asian languages in real time: speech-to-text, neural machine translation, text-to-speech, and LLM fine-tuning with RAG. My first LLM development work.

LLMSTTNMTTTSRAG

Diffusion-based demosaicing and compression

Ph.D. thesis · 2021 – 2024

Image demosaicing and compression with a denoising diffusion null-space model, including a learned color-filter-array pattern and a CFA optimization scheme for jointly optimal image quality and compression rate.

DiffusionDemosaicingCFAImage restoration

GAN super-resolution with feature-map adversarial loss

M.Sc. thesis · 2019 – 2021

A training scheme that applies the adversarial loss on the feature maps of the generated high-resolution image via pre-trained networks; outperformed the state-of-the-art CVPR method on a real-world low-resolution image benchmark at the time.

GANSuper-resolutionImage restoration

DeepView multi-plane view synthesis for ETRI

Dongseo · commissioned project

Reviewed and reimplemented DeepView: view synthesis with learned gradient descent over multi-plane images for ETRI, with reports and presentations to the institute.

View synthesisMPIPyTorch

Experience

2025 – present
AI Research Engineer, Laputa, Inc. (AI subsidiary of Sapyoung Publisher), Seoul
LLM research, the RAG and agentic workflow around a 20B+ open-weight multimodal MoE model, database, and cloud-GPU infrastructure; lead of the document-hierarchy repair project.
2024 – 2025
AI Research Engineer, STANS, Inc. (MDS Intelligence subsidiary), Seoul
Computer-vision models for intelligent CCTV: detection, crowd counting, violence detection; auto-labeling pipelines; quantization and deployment.
2019 – 2024
Graduate Student Researcher, Image Processing Lab, Dongseo University, Busan
Advisor: Prof. Suk-Ho Lee. Generative models for image restoration; five first-author papers; funded by NRF, ETRI, and Dongseo grants.
2019
Android App Developer, PT. Mat Ali Teknologi, Surabaya
Android chat application (Kotlin, Node.js).

Education

2021 – 2024
Ph.D., Computer Engineering, Dongseo University, Busan
Thesis: Image Compression and Demosaicing based on Denoising Diffusion Null-Space Model. GPA 4.33/4.50.
2019 – 2021
M.Sc., Computer Engineering, Dongseo University, Busan
Thesis: Super-Resolution with Adversarial Loss on the Feature Maps of the Generated High-Resolution Image. GPA 4.50/4.50.
2015 – 2019
B.Comp., Informatics Engineering, University of Surabaya
Thesis: Committee Management System for Dept. of Informatics in University of Surabaya. Summa cum laude, best graduate of the department. GPA 3.94/4.00.