AI researcher and engineer in Seoul, from Surabaya, Indonesia. I hold a Ph.D. in Computer Engineering and I work on vision–language models, currently on diagnosing and repairing fine-grained discrimination failures in Mixture-of-Experts VLMs; and on multimodal document understanding for retrieval systems. Before that: generative models for image restoration, and real-time computer-vision systems now running at Seoul subway stations.
I'm currently looking for a postdoctoral position (available from early 2027).
The Calibration Law: diagnosing and repairing answer-position bias in MoE vision–language models
Independent project · manuscript in preparation (ICML 2027 target) · project lead, with two collaborators
MoE VLMs 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, and establish, across eight open MoE VLMs, a "calibration law": training-free interventions succeed in proportion to how well they calibrate that bias. A label-free recipe greatly improves the VisMin group score and transfers zero-shot to Winoground and ColorSwap.
When can a second pass fix document structure? An end-to-end anatomy of hierarchy repair
Laputa, 2026 · manuscript in preparation (ICDAR 2027 target) · proposed and lead the project
PDF extractors recover text well and document hierarchy poorly. A self-supervised 115M pointer model, trained on pseudo-labels mined from section numbering and embedded PDF outlines with no human annotation, benchmarked end-to-end against deterministic rules, a 26B LLM, and the supervised 4B multimodal state of the art on three benchmarks. Preliminary results show deep, unnumbered book hierarchy is largely invisible to text-side and crop-level methods alike.
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
- Optimized Color Filter Array for Denoising Diffusion Null-Space Model-Based Demosaicing. I. Imanuel, H. Yang, S. Lee.
IEEE Access, 2024. ·SCIE - Denoising Diffusion Null-Space Model and Colorization based Image Compression. I. Imanuel, D. Kang, S. Lee.
Int. J. Internet, Broadcasting and Communication, 16(2), 2024 ·KCI - Demosaicing based Image Compression with Channel-wise Decoder. I. Imanuel, S. Lee.
Int. J. Internet, Broadcasting and Communication, 15(4), 2023 ·KCI - Image Compression with Channel-wise Decoder. I. Imanuel, S. Lee.
IEEE ICCE-Asia, 2022. ·🏅 Best Paper Award (Silver Prize) - Super-resolution with adversarial loss on the feature maps of the generated high-resolution image. I. Imanuel, S. Lee.
Electronics Letters, 58, 2022. ·SCIE
AI Research Engineer. Laputa, Inc. (AI subsidiary of Sapyoung Publisher), Seoul · Dec 2025 – present
- Develop an LLM-based multimodal Q&A system over Korean university textbooks
- Own the RAG and agentic workflow around a 26B VLM MoE model, the database and PDF-extraction pipeline, and cloud-GPU deployment (RunPod, Docker, EC2)
- Proposed and lead the document-hierarchy repair research project
LLM RAG agentic workflows vLLM Qdrant Unsloth Docker RunPod EC2
AI Research Engineer. STANS, Inc. (MDS Intelligence subsidiary), Seoul · Oct 2024 – Nov 2025
- Surveyed, fine-tuned, and quantized computer-vision models for intelligent CCTV: object detection, crowd counting, violence detection (action recognition + optical flow)
- Built auto-labeling pipelines and annotation guidelines; prepared the violence-detection system for KISA school-safety certification evaluation
- Detection and crowd-density systems deployed at Seoul subway stations
object detection action recognition crowd counting ONNX TensorRT quantization MLOps
Graduate Student Researcher. Image Processing Lab, Dongseo University, Busan · Sep 2019 – Aug 2024
- Generative models for image restoration under Prof. Suk-Ho Lee; five first-author papers
- Lab projects beyond the thesis: a commissioned DeepView implementation for ETRI, U-Net blood-vessel segmentation, CSRNet crowd counting
- Funded by NRF, ETRI, and Dongseo University research grants
GAN diffusion models image restoration view synthesis segmentation
Android App Developer. PT. Mat Ali Teknologi, Surabaya · Feb – Jul 2019
- Developed an Android chat application
Kotlin Android Node.js
- Ph.D., Computer Engineering. Dongseo University, 2021–2024. Thesis: Image Compression and Demosaicing based on Denoising Diffusion Null-Space Model. GPA 4.33/4.50.
- M.Sc., Computer Engineering. Dongseo University, 2019–2021. Thesis: Super-Resolution with Adversarial Loss on the Feature Maps of the Generated High-Resolution Image. GPA 4.50/4.50.
- B.Comp., Informatics Engineering. University of Surabaya, 2015–2019. 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.
- Unibook: Multimodal textbook Q&A. RAG and agentic workflow around a 26B VLM MoE model; vLLM, Qdrant, Unsloth, Docker, RunPod (Laputa)
- AWAS-Insight: Crowd-density and object detection for Seoul metro. Real-time crowd-counting and object-detection models, TensorRT/ONNX, deployed at Seoul subway stations (STANS)
- AWAS-Insight: School violence detection. Action-recognition and optical-flow models integrated with real-time object detector, prepared for KISA certification evaluation (STANS)
- Construction-site real-time translation. early-stage R&D: STT, NMT, TTS, LLM fine-tuning/RAG (STANS)
- Diffusion-based demosaicing and compression. learned CFA pattern and optimization scheme (Ph.D.)
- GAN super-resolution with feature-map adversarial loss. outperformed the SOTA CVPR method on a real-world benchmark at the time (M.Sc.)
- DeepView multi-plane view synthesis. commissioned reimplementation for ETRI (lab)
- Class projects: DQN on Sonic 2, adversarial masking of re-ID models, deep learning for network routing, genetic algorithms for TSP, curve interpolation from scratch
Full list with details → indraimanuel.github.io/projects
- Guest lecture, Vision–Language Models: Trends, Applications, and Implementation. Indonesian undergraduate webinar (ProjekinAja), Nov 2025
- Speaker, International Symposium on Management 2025 (University of Surabaya × Ho Chi Minh University of Banking). AI applications and data generation, May 2025
- Guest lecture for graduate students, University of Surabaya. Computer vision and image generation (CNNs, GANs, diffusion), Sep 2024
- Teaching assistant, Department of Informatics, University of Surabaya. Algorithms & Programming, Web Design, Web Programming, 2016–2018
Full list with PPTs (coming soon) → indraimanuel.github.io/talks
Frameworks & tools: PyTorch, TensorFlow, ONNX, TensorRT, OpenCV · vLLM, Unsloth, Qdrant, Hugging Face Transformers, bitsandbytes · Docker, RunPod, EC2 · MLOps and production model deployment
Languages: Indonesian (native) · English (IELTS 8.0) · Korean (KIIP Level 5 / TOPIK 4 equivalent)
- 🏝️ Field surveyor and translator for a Hansung University study on Indonesian and East Timorese laborers on remote Korean islands (Gaeya, Bogil, Nohwa), 2021
- 📱 Built a physics learning app for Indonesian high-school students for a volunteer education program (Unity/C#)
- 🎥 Indra in Korea. A YouTube channel about international student life and travel in Korea, 2019–2025; scripted, filmed, and edited everything myself
- 🏆 Excellent Researcher Award (Dongseo Graduate School) · Best Paper Award, IEEE ICCE-Asia 2022 · full scholarships across all three degrees
The story from Surabaya to Busan to Seoul → indraimanuel.github.io/about
