Ankit Sanjyal
I am an M.S. Data Science researcher at Fordham University, where I work as a Graduate Research Assistant in the Educational Data Mining Lab and as a Teaching Assistant for graduate Machine Learning courses.
My research focuses on flow and diffusion models, multimodal learning, and causal interpretability. I study how score-based and flow-matching generative models behave in structured latent spaces, asking questions like: how does energy evolve during a diffusion trajectory, when does classifier-free guidance hurt high-resolution quality, and can a causal graph over modalities make multimodal predictions explainable? My current work (Affect-Diff) couples a VAE over text/audio/video with a conditional 1D diffusion U-Net and a Gumbel-Softmax causal attention graph, targeting emotion recognition with counterfactual explainability. I have authored three preprints on arXiv covering energy dynamics in diffusion sampling (RectifiedHR), training-free local style control in SDXL (Local Prompt Adaptation), and a systematic negative-result study of DINO feature reuse for few-shot NeRF (NeRF Few-Shot).
Before Fordham, I was an Associate Data Consultant at Decision Foundry (Bangalore) and an undergraduate researcher at Jain University, where I worked on medical image segmentation and multiclass skin lesion classification. I am actively looking for research internships, gen ai/ ml research jobs and Ph.D. positions in generative modeling, multimodal AI, and representation learning.
M.S. Data Science, Fordham University · New York, NY
News
- 03/26 Actively working on Affect-Diff, a multimodal emotion recognition system with a causal attention graph and conditional 1D diffusion prior on CMU-MOSEI.
- 01/26 Teaching Machine Learning for Finance (Spring 2026) as TA at Fordham University.
- 07/25 Preprint Local Prompt Adaptation for Style-Consistent Generation on arXiv.
- 07/25 Preprint RectifiedHR: High-Resolution Diffusion via Energy Profiling on arXiv.
- 06/25 Preprint NeRF Few-Shot Limitations Analysis on arXiv.
- 08/25 Teaching Data Mining (Fall 2025) as TA at Fordham University.
- 08/24 Started as Graduate Research Assistant at Fordham University (MAESTRO project).
Publications
Selected Projects
-
Couples a causal attention graph (Gumbel-Softmax, T↔A↔V) with a VAE latent bottleneck and a conditional 1D diffusion U-Net (CFG + causal conditioning) for interpretable affect recognition on CMU-MOSEI. Counterfactual hallucination provides modality-level causal attribution.
-
Real-time cross-attention extraction from SD 2.1 with MAE < 10⁻⁶. CLIP embedding algebra for semantic steering (94% success). GPT-4o-mini chain-of-thought narratives. FastAPI + React 18 frontend.
-
Clean PyTorch reproduction of NoPropDT (Oxford, arXiv:2503.24322). Stack of local denoising blocks, each trained with per-layer MSE, zero gradient backward through the network. ~99% on MNIST, ~76% on CIFAR-10 without backpropagation.
-
Double Machine Learning (DML) pipeline for causal ATE/CATE estimation of momentum on next-day returns. Benchmarks: OLS, Random Forest, LSTM ablation (base/deep/wide). yfinance data for AAPL, MSFT, JPM, AMZN, XOM.
-
Full-stack platform (OpenAI script synthesis + PlayHT TTS + FFmpeg rendering) that auto-generates educational video lectures with retrieval-augmented QA bots. Live at Fordham University.
-
CrewAI multi-agent system (4 specialized agents) + Gemini that transforms raw data into interactive dashboards. 80% reduction in manual reporting time. Flask + React + PostgreSQL + Docker.
-
GPT-4o + MCP + MongoDB Atlas RAG autonomous agent that closes Jira tickets end-to-end: reads ticket, writes fix, opens PR. 60–90 second ticket-to-PR cycle. MUG Hackathon 2025.
-
Hybrid Neo4j (Dijkstra routing) + MongoDB + GPT-4o-mini natural language parsing for indoor campus navigation. React + Cytoscape.js interactive graph map.
Experience
| Aug 2025 – Present |
Teaching Assistant
Fordham University, New York, NY
Graduate courses: Data Mining (Fall 2025) and Machine Learning for Finance (Spring 2026). Weekly office hours, 1:1 research mentoring, and guest lectures.
|
| Aug 2024 – Present |
Graduate Research Assistant
Fordham University, Educational Data Mining Lab, New York, NY
Leading the MAESTRO project; exploring few-shot 2D-to-3D reconstruction on medical imaging data (OASIS dataset) using efficient neural representations.
|
| Jan 2023 – Jul 2024 |
Associate Data Consultant
Decision Foundry, Bangalore, India
Designed ETL workflows (DBT, Tableau Prep), optimized SQL on Amazon Redshift (40% faster retrieval on 1 TB+ datasets), delivered 25+ Tableau / Salesforce CRMA dashboards.
|
| Jan 2022 – Jul 2023 |
Undergraduate Research Assistant
Jain University, Bangalore, India
Deep learning pipelines for medical imaging; published work on toxic language detection and multiclass skin lesion classification (85% validation accuracy).
|
Education
| 2024 – Present |
M.S. Data Science
Fordham University, New York, NY
|
| 2019 – 2023 |
B.Tech Computer Science
Jain University, Bangalore, India
|