Senior Machine Learning Researcher
Jul 2026 — PresentLeading research and development of machine learning models for real-world applications in finance.
Senior ML Researcher @ RBC Borealis
I'm a Senior Machine Learning Researcher at RBC Borealis in Vancouver, where I work on LLM systems and foundation models for finance. I received my Ph.D. in Computer Science from Simon Fraser University (SFU) under the supervision of Prof. Yasutaka Furukawa, and my M.Sc. from Seoul National University (SNU) under Prof. Kyoung Mu Lee.
Latest
New preprint "Distilling the Shape of Uncertainty: What On-Policy Distillation Learns and How to Predict It" with SFU, on what on-policy distillation teaches LLMs and how to pick the right teacher.
Panelist on "Open Challenges in Grounding" at GroundLM 2026, the Workshop on Grounding Language Models at EMNLP 2026 in Budapest.
Promoted to Senior Machine Learning Researcher at RBC Borealis.
Gave a talk at Upper Bound 2026, Amii's annual AI conference in Edmonton.
Paper "FACTS: Table Summarization via Offline Template Generation with Agentic Workflows" accepted at ACL Findings 2026.
Paper "Embedding-based Context-aware Reranker" accepted at ICLR 2026.
1st place winner, Financial Agentic Retrieval Grand Challenge (FinAgentBench), AI for Finance Symposium '25.
Our team received the One RBC Team Award, recognizing groups whose collaboration brings the best of RBC to life through outstanding achievements.
Paper "Fast, Accurate, and Privacy-Compliant Table Summarization via Offline Template Generation" accepted at NeurIPS 2025 Workshop on Generative AI in Finance.
Defended my Ph.D. thesis on diffusion models for structured geometry at SFU.
Presented Visual Layout Composer at CVPR 2024, Seattle.
Joined RBC Borealis as a Machine Learning Researcher.
Paper "Visual Layout Composer" accepted at CVPR 2024.
Gave a talk on diffusion models for layout generation at Samsung AI Center, Toronto.
Paper "PuzzleFusion" accepted as Spotlight at NeurIPS 2023.
Paper "HouseDiffusion" accepted at CVPR 2023.
Paper "Scaleformer: Iterative Multi-scale Refining Transformers for Time Series Forecasting" accepted at ICLR 2023.
Career
Leading research and development of machine learning models for real-world applications in finance.
Contributing to research and development of machine learning models for real-world applications in finance.
EBCAR · ICLR 2026 ↗
FACTS · ACL Findings 2026 ↗
Table Summarization · NeurIPS Workshop 2025 ↗
Worked on a new method for automatically generating layout designs. Published at CVPR 2024.
Worked on Iterative Multi-scale Refining Transformers for Time Series Forecasting. Published at ICLR 2023.
Background
Research
Shows that on-policy distillation teaches a student not only how to branch (which tokens to consider) but where to branch (where to keep reasoning paths open), introduces the Entropy Shape Correlation (ESC) to measure the latter, and uses pre-distillation ESC to pick the best teacher.
On-policy distillation (OPD) is increasingly adopted in the post-training of large language models. However, a stronger teacher may fail to improve the student. Prior work attributes this to a lack of new knowledge, but what that knowledge is remains unclear. We show that OPD learns not only how to branch, i.e., which candidate tokens to consider, but also where to branch, i.e., where to keep reasoning paths open. The existing token-overlap metric captures only the former. We introduce the entropy shape, the sequence of next-token entropies, to characterize the latter, and propose the Entropy Shape Correlation (ESC) to measure student-teacher agreement in entropy shape. Tracking ESC throughout training and intervening on a teacher's entropy while preserving its token preferences, we empirically show that where to branch causally drives OPD gains, independently of how to branch. Finally, we formulate OPD outcome prediction as a teacher selection task and establish an evaluation protocol for it, under which pre-OPD ESC selects the best teacher for both students and outperforms existing predictors.
Introduces a framework for financial table summarization that combines offline template generation with agentic workflows, enabling structured and accurate natural language summaries of complex tabular financial data.
Query-focused table summarization requires generating natural language summaries of tabular data conditioned on a user query, enabling users to access insights beyond fact retrieval. Existing approaches face key limitations: table-to-text models require costly fine-tuning and struggle with complex reasoning, prompt-based LLM methods suffer from token-limit and efficiency issues while exposing sensitive data, and prior agentic pipelines often rely on decomposition, planning, or manual templates that lack robustness and scalability. To mitigate these issues, we introduce an agentic workflow, FACTS, a Fast, Accurate, and Privacy-Compliant Table Summarization approach via Offline Template Generation. FACTS produces offline templates, consisting of SQL queries and Jinja2 templates, which can be rendered into natural language summaries and are reusable across multiple tables sharing the same schema. It enables fast summarization through reusable offline templates, accurate outputs with executable SQL queries, and privacy compliance by sending only table schemas to LLMs. Evaluations on widely-used benchmarks show that FACTS consistently outperforms baseline methods, establishing it as a practical solution for real-world query-focused table summarization.
Proposes an embedding-based context-aware reranking method for retrieval-augmented generation (RAG) that improves document ranking by incorporating query context, leading to more relevant and accurate retrieval results.
Retrieval-Augmented Generation (RAG) systems rely on retrieving relevant evidence from a corpus to support downstream generation. The common practice of splitting a long document into multiple shorter passages enables finer-grained and targeted information retrieval. However, it also introduces challenges when a correct retrieval would require inference across passages, such as resolving coreference, disambiguating entities, and aggregating evidence scattered across multiple sources. Many state-of-the-art (SOTA) reranking methods, despite utilizing powerful large pretrained language models with potentially high inference costs, still neglect the aforementioned challenges. Therefore, we propose Embedding-Based Context-Aware Reranker (EBCAR), a lightweight reranking framework operating directly on embeddings of retrieved passages with enhanced cross-passage understandings through the structural information of the passages and a hybrid attention mechanism, which captures both high-level interactions across documents and low-level relationships within each document. We evaluate EBCAR against SOTA rerankers on the ConTEB benchmark, demonstrating its effectiveness for information retrieval requiring cross-passage inference and its advantages in both accuracy and efficiency.
Earlier workshop version of FACTS (ACL Findings 2026), presented at the NeurIPS 2025 Workshop on Generative AI in Finance.
Presents a dual diffusion model that jointly generates raster image content and corresponding vector layout for graphic design templates, enabling coherent, high-quality, and editable design generation from user-provided conditions.
This paper proposes an image-vector dual diffusion model for generative layout design. Distinct from prior efforts that mostly ignore element-level visual information, our approach integrates the power of a pre-trained large image diffusion model to guide layout composition in a vector diffusion model by providing enhanced salient region understanding and high-level inter-element relationship reasoning. Our proposed model simultaneously operates in two domains: it generates the overall design appearance in the image domain while optimizing the size and position of each design element in the vector domain. The proposed method achieves the state-of-the-art results on several datasets and enables new layout design applications.
Reformulates spatial jigsaw puzzle solving as a denoising diffusion process, jointly predicting piece positions and orientations without hand-crafted features or piece ordering assumptions, achieving state-of-the-art results on challenging benchmarks.
This paper presents an end-to-end neural architecture based on Diffusion Models for spatial puzzle solving, particularly jigsaw puzzle and room arrangement tasks. In the latter task, for instance, the proposed system "PuzzleFusion" takes a set of room layouts as polygonal curves in the top-down view and aligns the room layout pieces by estimating their 2D translations and rotations, akin to solving the jigsaw puzzle of room layouts. A surprising discovery of the paper is that the simple use of a Diffusion Model effectively solves these challenging spatial puzzle tasks as a conditional generation process. To enable learning of an end-to-end neural system, the paper introduces new datasets with ground-truth arrangements: 1) 2D Voronoi jigsaw dataset, a synthetic one where pieces are generated by Voronoi diagram of 2D pointset; and 2) MagicPlan dataset, a real one offered by MagicPlan from its production pipeline, where pieces are room layouts constructed by augmented reality App by real-estate consumers. The qualitative and quantitative evaluations demonstrate that our approach outperforms the competing methods by significant margins in all the tasks.
Generates structured vector floorplans using a diffusion model that simultaneously denoises discrete room types and continuous polygon coordinates, producing diverse and realistic architectural layouts from user-specified room programs.
The paper presents a novel approach for vector-floorplan generation via a diffusion model, which denoises 2D coordinates of room/door corners with two inference objectives: 1) a single-step noise as the continuous quantity to precisely invert the continuous forward process; and 2) the final 2D coordinate as the discrete quantity to establish geometric incident relationships such as parallelism, orthogonality, and corner-sharing. Our task is graph-conditioned floorplan generation, a common workflow in floorplan design. We represent a floorplan as 1D polygonal loops, each of which corresponds to a room or a door. Our diffusion model employs a Transformer architecture at the core, which controls the attention masks based on the input graph-constraint and directly generates vector-graphics floorplans via a discrete and continuous denoising process. We have evaluated our approach on RPLAN dataset. The proposed approach makes significant improvements in all the metrics against the state-of-the-art with significant margins, while being capable of generating non-Manhattan structures and controlling the exact number of corners per room.
Proposes an iterative multi-scale refining framework that applies Transformer-based forecasters at multiple temporal resolutions, progressively sharpening predictions from coarse to fine scales for improved time series forecasting accuracy.
The performance of time series forecasting has recently been greatly improved by the introduction of transformers. In this paper, we propose a general multi-scale framework that can be applied to the state-of-the-art transformer-based time series forecasting models (FEDformer, Autoformer, etc.). By iteratively refining a forecasted time series at multiple scales with shared weights, introducing architecture adaptations, and a specially-designed normalization scheme, we are able to achieve significant performance improvements, from 5.5% to 38.5% across datasets and transformer architectures, with minimal additional computational overhead. Via detailed ablation studies, we demonstrate the effectiveness of each of our contributions across the architecture and methodology. Furthermore, our experiments on various public datasets demonstrate that the proposed improvements outperform their corresponding baseline counterparts.
Converts raster building blueprint images into structured vector representations by detecting and reconstructing walls, doors, windows, and symbols as scalable primitives, enabling downstream architectural analysis.
This paper proposes a novel vectorization algorithm for high-definition floorplans with construction-level intricate architectural details, namely a blueprint. A state-of-the-art floorplan vectorization algorithm starts by detecting corners, whose process does not scale to high-definition floorplans with thin interior walls, small door frames, and long exterior walls. Our approach 1) obtains rough semantic segmentation by running off-the-shelf segmentation algorithms; 2) learning to infer missing smaller architectural components; 3) adding the missing components by a refinement generative adversarial network; and 4) simplifying the segmentation boundaries by heuristics. We have created a vectorized blueprint database consisting of 200 production scanned blueprint images. Qualitative and quantitative evaluations demonstrate the effectiveness of the approach, making significant boost in standard vectorization metrics over the current state-of-the-art and baseline methods.
Estimates 3D camera configurations for indoor panoramic images that have no visual feature overlap, leveraging structural layout constraints and Manhattan-world assumptions to recover global scene geometry.
This paper proposes an extreme Structure from Motion (SfM) algorithm for residential indoor panoramas that have little to no visual overlaps. Only a single panorama is present in a room for many cases, making the task infeasible for existing SfM algorithms. Our idea is to learn to evaluate the realism of room/door/window arrangements in the top-down semantic space. After using heuristics to enumerate possible arrangements based on door detections, we evaluate their realism scores, pick the most realistic arrangement, and return the corresponding camera poses. We evaluate the proposed approach on a dataset of 1029 panorama images with 286 houses. Our qualitative and quantitative evaluations show that an existing SfM approach completely fails for most of the houses. The proposed approach achieves the mean positional error of less than 1.0 meter for 47% of the houses and even 78% when considering the top five reconstructions.
Recovers audio signals from silent video by analyzing subtle, high-frequency vibrations in local image regions, improving over global visual microphone methods by capturing spatially-varying sound sources more accurately.
Sound waves cause small vibrations in nearby objects. A few techniques exist in the literature that can extract sound from video. In this paper we study local vibration patterns at different image locations. We show that different locations in the image vibrate differently. We carefully aggregate local vibrations and produce a sound quality that improves state-of-the-art. We show that local vibrations could have a time delay because sound waves take time to travel through the air. We use this phenomenon to estimate sound direction. We also present a novel algorithm that speeds up sound extraction by two to three orders of magnitude and reaches real-time performance in a 20KHz video.
Introduces a layer-wise progressive knowledge distillation strategy that transfers knowledge from a teacher to a student network one layer at a time, improving the efficiency and effectiveness of model compression.
Proves necessary and sufficient conditions for decomposing complete tripartite graphs into 5-cycles, extending known results on cycle decompositions in combinatorial graph theory.
The problem of finding necessary and sufficient conditions to decompose a complete tripartite graph Kr,s,t into 5-cycles was first considered by E.S. Mahmoodian and Maryam Mirzakhani (1995). They stated some necessary conditions and conjectured that those conditions are also sufficient. Since then, many cases of the problem have been solved by various authors; however, the case when the partite sets r≤s≤t have odd and distinct sizes remained open. We show the conjecture is true when r, s and t are all multiples of 5, t+90 ≤ 4rs/(r+s), and t ≠ s+10.
Characterizes uniquely k-list colorable planar graphs and graphs on surfaces, establishing new structural properties and bounds for list coloring in graph theory.
A graph G is called uniquely k-list colorable (UkLC) if there exists a list of colors on its vertices, say L={Sv | v ∈ V(G)}, each of size k, such that there is a unique proper list coloring of G from this list of colors. A graph G is said to have property M(k) if it is not uniquely k-list colorable. Mahmoodian and Mahdian characterized all graphs with property M(2). For k≥3 property M(k) has been studied only for multipartite graphs. Here we find bounds on M(k) for graphs embedded on surfaces, and obtain new results on planar graphs. We begin a general study of bounds on M(k) for regular graphs, as well as for graphs with varying list sizes.
IP
Describes hierarchical retrieval over long documents: passage-level embeddings are aggregated by a higher-level encoder into document embeddings, so a query first locates the most relevant documents and then the most relevant passages within them.
Describes a diffusion framework that jointly generates design layouts in image and vector domains, using visual element characteristics as conditions to improve diversity and saliency reasoning.
Describes an iterative multi-scale neural network approach for time series forecasting that progressively refines predictions from coarse to fine temporal resolutions, achieving 5.5–38.5% error reduction across benchmarks.
Community
Fun
In PuzzleFusion (NeurIPS 2023) we showed that a diffusion model can solve spatial jigsaw puzzles: treat the scattered pieces as noise, then denoise them into place. Here is the human benchmark — tap two tiles to swap them and reassemble the photo. Or press the button and let reverse diffusion do it for you.
In Scaleformer (ICLR 2023) we improved time-series forecasting by refining predictions from coarse to fine scales. Your turn: study the history, then draw your forecast across the shaded region with your mouse or finger. Reveal the future to see whether you beat the model.
history you model truth
Contact
Feel free to reach out about research, collaborations, or anything interesting.
aminshabaany@gmail.comVancouver, Canada