Hi! I'm Deb.
My name is Debjyoti Mondal. I'm a Research and Development Engineer at Samsung.
And I like a lot of things.
Work Experience
Lead Research Engineer, Natural Language Intelligence
Studying non-euclidean ML and Graph Neural Networks.
Designed the Device Assist QA feature introduced in Galaxy S26, and scaled it to 10M+ queries/day using efficient retrieval and LangChain. I had incredible fun building this! 😁
Enabling multi-objective alignment in Language Models with focus on Safety and Helpfulness, any model can handle multiple objectives with < 1% drop on general benchmarks.
Research Engineer, Natural Language Understanding
Enabled multi-modal reasoning with GNNs, published at AAAI 2024 — smallest model beat the baseline by ≥10%, new SoTA on ScienceQA.
Developed an LLM safety module for Bixby and GalaxyAI pipelines (on-device + cloud), handling 12 locales and 9 languages, scaled to 10M+ calls/day.
Maintained key modules in the Bixby pipeline, improving production performance to 95%.
How are we doing? Can't say much, but check this out.
Student Trainee 📜
Worked on the interpretability of Language Models for Intent Classification and Slot Tagging tasks.
Used Language Interpretability Tool (LIT) and Local Interpretable Model-agnostic Explanations (LIME).
Machine Learning Intern 📜
I developed a generic object detection and localization framework.
And also some OCR apps for template based info extraction.
Hehe, this was fun - implemented deskew and denoise for scans using Hough transform and Thresholding.
Education
Bachelor of Technology in Electronics and Communication Engineering
Publications
SmoGVLM: A Small, Graph-Enhanced Vision-Language Model 📄
A super-fast method to inject knowledge into Small Language Models!
ICASSP 2026Can Safety Emerge from Weak Supervision? A Systematic Analysis of Small Language Models 📄
We devise a pipeline for aligning an LLM with as many objectives as you want.
**With < 1% performance degradation on general tasks.**
We show it's effectiveness on Small Models with a focus on Safety and Helpfulness.
RG-VQA: Leveraging Retriever-Generator Pipelines for Knowledge Intensive Visual Question Answering 📄
Work done in collaboration with IIT-Bombay.
EMNLP Findings 2025From Perception to Reasoning: Enhancing Vision-Language Models for Mobile UI Understanding 📄
Work done in collaboration with IIT-Bombay. We setup a new benchmark with some really complex queries, that need screen understanding and UI grounding.
ACL Findings 2025KAM-CoT: Knowledge Augmented Multimodal Chain-of-Thoughts Reasoning 📄
We introduce knowledge graphs as a new modality, keeping the graph structure intact.
With 2 graph convolutional layers as the KGEncoder, our method gains a deeper contextual understanding of text, therefore reducing hallucinations and enhancing the quality of answers.
Seg-HGNN: Unsupervised and Light-Weight Image Segmentation using Hyperbolic GNNs 📄
We exploit the properties of Hyperbolic Spaces to do segmentation tasks in very low dimensions.
With the extracted image features, we build a graph in the Lorentz space, and perform unsupervised clustering to get semantically similar regions.
Beyond Work
Some cool projects
Modeling Chaotic Epidemic Models on FPGAs
A general procedure to implement Chaotic Models. Used Euler's Forward method to discretize continuous-time differential equations, then generated RTL Schematics and compared the resources used. Ahaa! And using this, I studied period-doubling bifurcations & got to the Feigenbaum constant with the logistic map and the Genesio-Tesi attractor. Be sure to watch this. It's pretty.
Bachelor's ThesisFraJuVis - Julia Fractals on Android
An interactive Android app for visualizing Julia fractals in real time. Tap anywhere on the screen to set the complex parameter c in p(z) = z² + c, dynamically generating fractal patterns. Efficiently computes and renders using recursive iteration.
AndroidDeep Dream
Implementation of the DeepDream algorithm. Extracts features from InceptionV3, and using gradient ascent over iterations, enhances 😵💫 patterns in the input image.