Janmenjaya Panda

Researcher at Adobe | GSoC ’25 – ’24 at SageMath | IIT Madras ’24

Research Interests: Computer Vision, Multi-Modal Large Language/ Reasoning Models, Generative AI, Graph theory, Algorithms

I am an AI researcher interested in machine learning, computer vision, and vision–language models, with a focus on representation learning, multimodal reasoning, and evaluation in latent spaces. My background includes graphs, algorithms, and mathematical modeling, which informs my approach to building principled and scalable learning systems.

Bachelor of Technology

Indian Institute of Technology, Madras

• Mechanical Engineering (Major)
• Artificial Intelligence & Computing (Minor)

2020-2024

CGPA: 08.88/10.00

Graduation Month: 07/2024

• Math, Probability & Finance:
  Differential Equations (MA2020), Probability, Statistics and Stochastic Processes (MA2040), Stochastic Calculus for Finance (MA5018), Economic Network Analysis (HS3100)
• Computation:
  Stochastic Processes in Mechanics (AM5340), Finite Element Analysis (ME5204)
• Systems:
  Database Management System (NPTEL), Computer Networks and Internet Protocol (NPTEL)
• Optimization & Algorithms:
  Nonlinear Optimization: Theory and Algorithms (CS5020), Linear Programming and Combinatorial Optimization (CS5210), Advanced Graph Algorithms (CS6130), Approximation Algorithms (CS6841), Design and Analysis of Algorithms (NPTEL), Advanced Graph Theory (NPTEL)
• Machine Learning:
  Pattern Recognition and Machine Learning (CS5691), Multi-armed Bandits (CS6046), Bandits Algorithms (NPTEL), Natural Language Processing (CS6370), Reinforcement Learning (CS6700), Fundamentals of Deep Learning (CS6910)
• Sports:
  Chess (NSO)

Class XII

Tetrahedron Higher Secondary School, Tangi, Odisha

2018 - 2020

Percentage: 90.67%

Council of Higher Secondary Education, Odisha

Physics, Chemistry, Mathematics, Biology, Odia, English

Class X

Saraswati Shishu Vidya Mandir, Paradeep, Odisha

2012 - 2018

Percentage: 96.83%

English, Odia, Hindi,Mathematics, Science, Social Studies

Board of Secondary Education, Odisha

• Activities and societies: Actively participated in inter-school and inter-district competitions in essay writing, drawing, science, and cultural quizzes.Activities and societies: Actively participated in inter-school and inter-district competitions in essay writing, drawing, science, and cultural quizzes.
• Secured Board #2 in Secondary Education Certificate Examination 2018, Board of Secondary Education, Odisha.

Member of Technical Staff

Illustrator, Adobe

Bengaluru, Karnataka, India

May 2024 - Present

• Scalable Vision-Language Framework for Intent Preserving Layout Adaptation with Retrieval-Guided Vision Prior (Under review at CVPR 2026)
• Investigated how to preserve design intent and visual hierarchy during aspect-ratio–conditioned layout adaptation.
• Developed a novel retrieval-guided synthetic weak supervision generation pipeline and vision–language-conditioned adaptation framework to generate design layout variants that preserve visual hierarchy and design intent across diverse aspect ratios.
• Introduced novel evaluation metrics to quantify preservation of relative visual importance and semantic grouping of design elements, demonstrating a 21% and 41% improvement respectively over state-of-the-art layout adaptation methods.
• Natural-Language Guided Object Addition thru Contextual Reasoning
  (Invention disclosure approved for U.S. patent filing; USPTO submission in preparation)
• Developed a fully automated, context-aware pipeline for identity-preserving object addition from natural language prompts, finetuning Florence-2-Large to accurately predict spatial insertion points in complex scenes, furthermore finetuning FLUX.1-dev, to insert an object from image without object, prompt, and the predicted spatial point of object insertion.
• Designed a novel data synthesis framework using multiple multimodal large language models and reasoning-based grounding models to generate diverse training instances, improving spatial accuracy, prompt compliance, and visual fidelity without requiring object masks, reference images, or manual annotations.
• Consistent View Generation of Vector Graphics under Specified Azimuth and Elevation
  (Data generation)
• Curated and filtered open-source 3D assets from TexVerse and TurboSquid under permissible licenses, selecting aesthetic assets and performing multi-variant non-photorealistic rendering (NPR) to generate consistent multi-view vector sets across diverse viewpoints.
• Constructed an automated viewpoint annotation pipeline by finetuning PoseDiffusion on synthesized multi-view renderings, enabling azimuth and elevation estimation for unannotated relevant Adobe Stock assets.
• Natural-Language Guided Layered Vector Graphics Generation
  (Data generation)
• Generated a dataset of multi-layered vector graphics from Adobe Stock images, using semantic segmentation, occlusion classification, inpainting, and heuristic-based layering, followed by refinement via chain-of-thought reasoning on a vision–language model.

Open-Source Contributor, SageMath

Google Summer of Code, 2025

May 2025 - September 2025

• Project: On the Implementation of the 𝓞(|E|•√|V|) Micali-Vazirani Algorithm
• Mentors: Prof. David Coudert, Prof. Dima Pasechnik
• Abstract: Matchings and perfect matchings have received considerable attention in graph theory as well as in other related domains (such as, but not limited to, algorithms and optimization). There still remain many open problems — such as Barnette’s conjecture, Berge-Fulkerson conjecture, and so on — due to which it continues to remain an active area of research. At the heart of all this research lies the bottleneck of finding a maximum cardinality matching in undirected graphs. The objective of this project is to implement the Micali-Vazirani algorithm in SageMath — which achieves the best known theoretical runtime of 𝓞(|E|•√|V|) for computing such a matching — and to make all of this available freely to students, educators as well as researchers all across the world.
• Proposal: Link
• Verification: Link
• Completion: Link
• Documentation: Link

Open-Source Contributor, SageMath

Google Summer of Code, 2024

May 2025 - September 2025

• Project: On Decompositions, Generation Methods and related concepts in the theory of Matching Covered Graphs
• Mentor: Prof. David Coudert
• Abstract: Matchings and perfect matchings have received considerable attention in graph theory as well as in other related domains (such as, but not limited to, algorithms and optimization). There still remain many open problems — such as Barnette’s conjecture, Berge-Fulkerson conjecture, and so on — due to which it continues to remain an active area of research. For problems concerning perfect matchings, it is well-known that it suffices to solve them for matching covered graphs (that is, those connected graphs wherein each edge belongs to some perfect matching). The objective of this project is to implement efficient algorithms pertaining to the canonical partition, tight cut decomposition, dependency relations, (optimal) ear decomposition, brick and brace generation methods and related concepts in the theory of matching covered graphs, and to make all of these available freely to students, educators as well as researchers all across the world.
• Proposal: Link
• Verification: Link
• Completion: Link
• Documentation: Link
• SageMath Page: Link

Product Intern

Adobe Content Platform Collaboration, Adobe

Bengaluru, Karnataka, India

May 2023 - July 2023

• On the Automation of Publishing of Content Delta
• Proposed and implemented a Java client SDK to automate the publishing of predefined and custom content deltas from Adobe Content Platform clients to Adobe assets via Live Edit Services (LES), in collaboration with the Copy-Archive-Move-Purge team.
• Engineered protocol buffer–based delta serialization, real-time publishing over WebSocket, and automated integration tests for copy, move, and rendering workflows to ensure correctness of live asset updates.

Young Research Fellow

Reseach In Algorithms and Graphs Lab, IIT Madras

September 2022 - April 2023

• Birkhoff-von Neumann graphs that are PM compact: An alternative proof
• Guide: Prof. Nishad Kothari
• \item Characterized matching covered graphs wherein \textbf{each removable edge leads to a wheel graph}. \item Developed a simplified proof for the 𝓟-characterization of Birkhoff-von Neumann graphs that are perfect matching compact, as an application of the previous characterization.
• Poster: Link

Coordinator, Students' Website

Institute WebOps and MobOps, IIT Madras

August 2022 - April 2023

• Contributed to the enhancement and maintenance of multiple dynamic modules on the IIT Madras students’ website.
• Designed and developed the Academic section of the students’ portal, featuring course search, structured overviews, and an integrated feedback system.

Project Member, Visualizer Library

Programming Club, Center for Innovation, IIT Madras

May 2021 - Apr 2022

• Contributed to the development of the Visualizer Library, an online tool designed to illustrate real-world graph algorithms in practice.
• Implemented visualizations for graph algorithms, including Dinic's Max-Flow Algorithm for Splitwise and the Bipartite Stable Marriage Problem, enabling practical understanding and analysis.
• Developed an interactive website designed to visually showcase manifold graph-pathfinding algorithms, including DFS, BFS, A* search, Greedy Best-First Search, Swarm Search, Bidirectional Swarm Search, and Convergent Swarm Search.

Project Member, Software Module, Triphibian

ibot Club, Center for Innovation, IIT Madras

May 2021 - Apr 2022

• Contributed to the development of the automation module for Triphibian, an autonomous all-terrain vehicle designed for surveillance applications.
• Integrated sensor systems, including Inertial Measurement Unit (IMU), pressure, and humidity sensors, with a Raspberry Pi for data acquisition and processing.
• Designed and implemented a robust PID control system featuring anti-windup mechanisms, noise filtering, and parameter tuning to optimize the robot's performance.

ICPC 2023: All India Rank 102 (Team: Mech20Coderz) in ICPC Amritapuri Qualifier Round 2023; All India Rank 436 (Team: Mech20Coderz) in ICPC Kanpur-Mathura Qualifier Round 2023

YRF 2022: Distinguished as one of the 32 students selected as Young Research Fellow in the batch of 2022 in IIT Madras

Mimamsa 2021: Second runner-up in National level Science-Quiz Competition Mimamsa 2021, organized by IISER Pune.

NEST 2020: All India Rank 101 in NEST 2020 conducted by NISER, Bhubaneswar, Odisha.

JEE Advanced 2020: All India Rank 2300 in JEE Advanced 2020 conducted by IIT Delhi.

Programming: Python, C/C++, Java, Bash

Machine Learning: Pytorch, Tensorflow, OpenCV

Algorithms & Maths: SageMath, MATLAB

Systems & Tools: Git/GitHub, Docker

On Options Framework and Hierarchical Reinforcement Learning

Course Project: Reinforcement Learning (CS6700)

January 2024 - May 2024

• Collaborator: Nishant Sahoo
• Description:
  • Implemented and compared two reinforcement learning algorithms, SMDP (Semi-Markov Decision Process) Q-Learning and Intra-Option Q-Learning, on the Taxicab Domain environment. The comparison was based on the optimal policy learned, the rate and quality of convergence, and an analysis against an analogous non-option framework.
  • Proposed situationally better options that may lead to quicker convergence to the optimal policy compared to the default options provided.
• Implementation & Report: Link

On Deep and Duelling Q-Network and Monte-Carlo REINFORCE

Course Project: Reinforcement Learning (CS6700)

January 2024 - May 2024

• Collaborator: Nishant Sahoo
• Description:
  • Implemented and compared different variants of Dueling-DQN (one with average advantage and the other with max advantage) and Monte-Carlo REINFORCE (with and without baseline) algorithms across two Gymnasium environments: Acrobot-v1 and CartPole-v1.
  • Observed several characteristics, such as:
    • How decoupling the state value function from the action value function by introducing advantage results in better performance.
    • How the introduction of a baseline reduces the variance of the estimated return, introduces a bias to the estimation, achieves faster convergence, improves the exploration-exploitation trade-off, and stabilizes the learning process by preventing large fluctuations in the estimated returns in the REINFORCE algorithm.
• Implementation & Report: Link

On Information Retrieval System

Course Project: Natural Language Processing (CS6370)

January 2024 - May 2024

• Collaborator: Nishant Sahoo, Rahul Verma
• Description:
  • Enhanced traditional retrieval systems by integrating advanced techniques such as term frequency-inverse document frequency (TF-IDF), Latent Semantic Analysis (LSA) for relevant concepts and context-aware embeddings, and Latent Dirichlet Allocation (LDA) for topic extraction.
  • Incorporated state-of-the-art embeddings like BERT and word2vec to capture nuanced contextual relationships, alongside WordNet for similarity propagation and LSTM networks for query auto-completion.
  • Evaluated and compared system performance of both a basic classical IR system and one using deep learning models, using metrics like Precision, Recall, F0.5 Score, and nDCG on the Cranfield dataset.
• Implementation & Report: Link

On Image Captioning

Course Project: Fundamentals of Deep Learning (CS6910)

January 2024 - May 2024

• Collaborator: Nishant Sahoo
• Description:
  • Implemented and compared image captioning models using a CNN with NetVLAD as an encoder, and single-layer RNN and LSTM decoders, analyzing their performance on caption generation tasks.
  • Achieved better performance with the LSTM decoder, demonstrating lower losses and higher BLEU scores (1, 2, 3, 4), due to its ability to capture long-range dependencies and effectively address the vanishing gradient problem.
• Implementation & Report: Link

On Machine Translation

Course Project: Fundamentals of Deep Learning (CS6910)

January 2024 - May 2024

• Collaborator: Nishant Sahoo
• Description:
  • Developed a machine translation system using an encoder-decoder architecture with single-layer LSTM networks, leveraging GloVe embeddings for English and IndicBERT embeddings for Indian languages to address diverse linguistic characteristics.
  • Evaluated translation performance using BLEU@k scores (k=1, 2, 3, 4), ensuring a comprehensive assessment of translation quality across multiple $n$-gram levels.
• Implementation & Report: Link

On the Detection of Spam Emails

Course Project: Pattern Recognition & Machine Learning (CS6370)

January 2024 - May 2024

• Description:
  • Designed and implemented a spam email classification system using various machine learning methods, including Logistic Regression, Naive Bayes variants (Gaussian, Multinomial, Bernoulli), Support Vector Machines (SVM), and Decision Trees.
  • Enhanced model performance through ensemble methods like boosting and bagging, including Random Forest, XGBoost, and LightGBM, while optimizing feature extraction with TF-IDF, $n$-grams, and metadata encoding.
  • Developed a machine translation system using an encoder-decoder architecture with single-layer LSTM networks, leveraging GloVe embeddings for English and IndicBERT embeddings for Indian languages to address diverse linguistic characteristics.
• Implementation & Report: Link

AI for Sustainable Finance

Hackathon: Encryptcon 2024

January 2024

• Collaborator: Balakumar Ravindran, Guru Shreyaas, Mohan Kumar
• Description:
  • Developed a model to predict trends and risks in green finance, enabling informed and sustainable investment decisions.
  • Designed an optimal stock allocation strategy by minimizing the green index using a transformer-based architecture that integrates financial and sustainability data through cross-attention between stock price sequence encoding and transcript encoding for sentiment analysis.
• Implementation & Report: Link

On Spatial Patterns in Epidemic Modelling due to Cross-Diffusion Terms

Course Project: Finite Element Analysis (ME5204)

July 2023 - November 2023

• Collaborator: Nishant Sahoo
• Description:
  • Performed a numerical analysis of a two-dimensional transient epidemic model with nonlinear self-diffusion and cross-diffusion terms, capturing the spatial dynamics of susceptible and infected populations.
  • Discretized the reaction-diffusion system using the finite element method and simulated both forward and backward Euler schemes, approximating the nonlinear systems using Picard’s iteration and the Newton-Raphson method.
  • Investigated the influence of several parameters, such as birth rate, carrying capacity, disease transmission rate, and recovery rate; highlighted the impact of diffusion terms on emergent complex spatial structures like stripes, spots, and holes, influenced by initial conditions.
  • Analyzed the stability and convergence of the numerical scheme both in time and space, providing relevant error estimates.
• Implementation & Report: Link

Distributed Exploration in Multi-Armed Bandits

Course Project: Multi-armed Bandits (CS6040)

July 2023 - November 2023

• Collaborator: Archish S
• Description:
  • Explored pure exploration in a Stochastic Multi-Armed Bandit setting where p players collaborate to identify an ε−optimal arm.
  • Demonstrated a parallel speedup by a factor of p in the fixed confidence setting and extended the hypothesis to a fixed budget setting, showing an analogous performance boost.
• Implementation & Report: Link

On Solving Differential Equations using Physics-Informed Neural Networks

Course Project: Machine Learning in Engineering and Scientific Applications (ID5030)

January 2023 - May 2023

• Description:
  • Studied and implemented Physics-Informed Neural Networks for solving partial differential equations and ordinary differential equations, such as the advection equation.
  • Implemented Lagaris' method from the paper Artificial Neural Networks for Solving Ordinary and Partial Differential Equations and the Raissi discrete-time model from Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations.
• Implementation & Report: Link

Biomedical Image Classification using U-Net Architecture

Course Project: Machine Learning in Engineering and Scientific Applications (ID5030)

January 2023 - May 2023

• Description:
  • Implemented U-Net, a convolutional network for biomedical image segmentation, to accurately segment the fetal head. Used Dice score as the loss metric and achieved an $F_1$ score > 0.9 on the test dataset, drawing insights from the paper `U-Net: Convolutional Networks for Biomedical Image Segmentation'.
• Implementation & Report: Link

Stochastic Modelling of Lake Eutrophication

Course Project: Stochastic Processes in Mechanics (AM5340)

January 2022 - May 2022

• Description:
  • Applied Monte Carlo simulations to analyze a bistable one-dimensional stochastic lake eutrophication process and compared it with the deterministic process (the latter solved numerically using the Runge–Kutta method).
  • Simulated time histories using Shinozuka's approach (which relies on the fact that a random process can be simulated by a series of cosine functions with random frequencies), assuming a suitable power spectral density (PSD) function, and compared the target probability density function (pdf) and target PSD.
  • Developed a surrogate model using polynomial chaos expansion to efficiently characterize and analyze the behavior of this highly nonlinear, noisy system.
• Implementation & Report: Link

Schelling's Model of Global Segregation Arising from Local Homophily

Course Project: Economic Network Analysis (HS3100)

January 2021 - May 2021

• Description:
  • Implemented and studied Schelling’s model of global patterns of spatial segregation, which evolve from the effect of homophily (a sociological tendency of individuals to associate and bond with similar others) operating at the local level.
  • Analyzed the impact of several parameters, such as homophily threshold, population size, and empty house ratio, on the convergence of the global mean similarity ratio.
• Implementation & Report: Link

On Conway's Game of Life

Project Focus: Emergent behavior in cellular automata

December 2023

• Description:
  • Implemented the Turing-complete, zero-player Game of Life using simple HTML, CSS, and JavaScript, depicting the unpredictability of simple systems.
  • Explored the emergence of complex patterns, such as gliders and still lifes, and observed the unpredictability of simple systems.
• Implementation & Report: Link

Solving Wordle using Information Theory

Project Focus: Information Theory

September 2023

• Description:
  • Built a ReactJS web app for solving Wordle, averaging 3.43 steps for completion among all possible five-letter words.
  • Employed an entropy-driven approach using the relative frequency of English letters for efficient word selection.
• Implementation & Report: Link

On Visualizing Several Pathfinding Algorithms on a Graph

Project Focus: Algorithms

September 2023

• Description:
  • Built a web application that provides a visual representation of various pathfinding algorithms such as Dijkstra's algorithm, A* Search (weighted), Greedy Best-first Search (weighted), Swarm Algorithm (weighted), Convergent Swarm Algorithm (weighted), Bidirectional Swarm Algorithm (weighted), Breadth-first Search (unweighted), and Depth-first Search (unweighted), allowing users to observe how different algorithms work and understand their efficiency.
  • Implemented several patterned maze generation methods such as recursive division (without any skew, with vertical skew, and with horizontal skew), random maze, basic weight maze, and stair patterns as some default obstacles examples.
• Implementation & Report: Link

On Visualizing Several Sorting Algorithms

Project Focus: Algorithms

August 2023

• Description:
  • Developed a web application that provides a visual representation of several sorting algorithms, such as bubble sort, selection sort, insertion sort, quick sort, and merge sort, allowing users to observe how different algorithms work and understand their sorting efficiency.
  • Each algorithm is implemented as an asynchronous function that takes an array of DOM element objects as a parameter and sorts these elements in-place, in ascending order.
  • The visualization is achieved using custom-documented APIs for rendering and color-coding, making it easier to follow the sorting process step by step.
• Implementation & Report: Link

Student Mentor

Alumni and Corporate Relations, IIT Madras

June 2024 - December 2024

• Guided and mentored a group of six final-year undergraduate students during the placement session.

Campus Coordinator

Mimamsa 2022, IISER Pune

December 2021 - May 2022

• Served as an ambassador for Mimamsa, organizing workshops, events, and outreach activities to promote the competition within our institute.

English: Professional working proficiency

Hindi: Native or bilingual proficiency

Odia: Native or bilingual proficiency