01
ML & Research
Federated and personalized learning, computer vision and LLMs — designed with ablations, measured honestly, and reproducible.
PyTorchMONAIFederated LearningLLMsCV
( Portfolio — 2026 )
ML Researcher & Full-Stack Engineer
teaching models to learn together, without sharing secrets.
✦ Federated & Decentralized Learning✦ Trustworthy ML✦ Distributed AI / LLM Systems✦ Full-Stack Web AppsScroll
(01)About
Iworkwheremachinelearningmeetsrealproducts.Itrainfederatedmodelsacrosshospitalsthatcan'tsharescans—andIbuildthefullstackaroundthem:FastAPIservices,Next.jsdashboards,andpipelinesthatturnamountainofgovernmentPDFsintodatayoucantrust.
M.S. Computer Science at the University of Florida. I research personalized federated learning, published first-author computer-vision work, and ship the full stack around models — APIs, databases and the interfaces people actually use.
0.00GPA at UF (/4.0)
0.0M+records wrangled at RBI
0%less FL comms overhead
0stRajasthan Police Hackathon
University of Florida
M.S. Computer Science
University of Mumbai
B.Tech AI & Machine Learning
Federated Learning✺Full-Stack Apps✺Computer Vision✺LLM Pipelines✺Distributed Systems✺Medical Imaging✺
(02)What I do
I don't stop at the notebook. I take models from experiment to API to the interface people actually click on.
02
Full-Stack Engineering
From schema to API to UI. Typed React/Next.js frontends on top of FastAPI/REST backends that stay fast over million-row datasets.
Next.jsReactTypeScriptFastAPIPostgreSQLMongoDB
03
Data & LLM Pipelines
Messy PDFs and spreadsheets in, clean queryable data out — with caching, dedup and provenance back to the source page.
DuckDBParquetPySparkPandasRedisNeo4j
(03)Experience
Jun — Aug 2026
Mumbai
- End-to-end agricultural intelligence pipeline ingesting government publications from 5+ sources into normalized, linked, queryable datasets.
- Hybrid extraction with Gemma 4 31B + Camelot: schema-constrained zero-shot extraction, NDJSON recovery, caching, dedup and source-level provenance.
- Scaled analytics to 1.2M+ records by moving pandas/CSV to Parquet + DuckDB.
- FastAPI + Next.js dashboard with incremental loading over million-row datasets.
Gemma 4DuckDBParquetFastAPINext.js
(04)Research
Personalized Federated Learning2026
Dual experts for multimodal 3D medical segmentation
Distributed ML Research · University of Florida
Can modality-specialized experts stop CT and MRI from fighting each other in non-IID federated training? A SwinUNETR with a globally shared encoder and private, modality-specific latent spaces and decoders — FedAvg only touches the 42.58M-param encoder while ~102M expert params never leave the client.
0.00CT Dice (federated)
0.000MRI Dice
0%fewer params exchanged
0FL rounds · 3 clients
01
First-author · Peer reviewed2026
GATE-Dehaze: a GAN-attention-transformer ensemble
Turkish Journal of Remote Sensing
An adaptive ensemble of four GAN-based U-Nets — vanilla, Transformer, Transformer + CBAM, Transformer + channel attention — fused to strip haze from satellite imagery while keeping local detail, long-range context and edges intact.
0.0000SSIM · moderate haze
0.0000SSIM · thin haze
0models in the ensemble
02
⟷
HazyGATE-Dehaze(05)Selected projects
Things
I've built
Keep scrolling → the page moves sideways.
01 / 05
Agri Intelligence
PDFs from five government sources in, one queryable 1.2M-row knowledge base out — with every number traceable to its source page.
LLM extractionFastAPINext.jsDuckDB
02 / 05
FedDual Experts
Shared encoder, private experts. Personalized FL for CT/MRI organ segmentation on AMOS22.
PyTorchMONAIFedAvg
Published · 2026
03 / 05
GATE-Dehaze
Four attention-flavoured GANs, one ensemble. Clear skies for satellite imagery.
GANsTransformersRemote Sensing
04 / 05
Sentinel Eye
Red-light & helmet violations, ANPR, camera-based pulse estimation, drone + face recognition for border security.
YOLOOCRFace Rec
Research · 2025
05 / 05
Satellite VQA
Ask questions about satellite images in plain English, plus a real-time statistical mosaicking pipeline.
VQAMosaicking
(06)Toolkit
Languages/Python/TypeScript/JavaScript/Java/SQL/
ML / Research/PyTorch/Federated Learning/LLMs/NLP/Computer Vision/MONAI/
Frontend/React/Next.js/TypeScript/JavaScript/
Backend/FastAPI/REST APIs/PostgreSQL/MongoDB/SQLite/Redis/
Data/PySpark/Apache Spark/DuckDB/Parquet/Pandas/Neo4j/
PythonTypeScriptJavaScriptJavaSQLPyTorchFederated LearningLLMsNLPComputer VisionMONAIReactNext.jsFastAPIREST APIsPostgreSQLMongoDBSQLiteRedisPySparkApache SparkDuckDBParquetPandasNeo4j
(07) — Contact