Sanket Deshpande
Download CV
MS Computer Science Georgia State University • 2023-2025
BS Computer Science Vellore Institute of Technology • 2016-2020

Sanket Deshpande

I build and ship AI products end-to-end — from LLMs and multimodal systems to distributed training and real-time inference. My edge: speed, ownership, and turning research into production.

Research to Production

PyTorch DDP
Custom Architectures
3D CNNs
Transformers
Multi-GPU Training
SLURM / HPC
Experiment Tracking

AI Product Engineering

LLM APIs (Gemini, GPT)
RAG Systems
Vector DBs
Multimodal AI
Prompt Engineering
Agent Systems
Streaming Inference

Full-Stack Development

Python
FastAPI
Postgres
React / Next.js
REST / WebSocket
FFmpeg
S3 / Cloud Storage

Deployment & Scale

Docker / Containers
AWS / GCP
CI/CD
Model Serving
Monitoring
Load Balancing
Cost Optimization

Professional Journey

Building impactful solutions across various industries

Founding ML Engineer

Digital Studio Labs • San Francisco, CA

Jul 2025 - Present
  • Cre8able: Built and launched a multimodal video-editing platform from idea to live beta in 6 weeks; cut creator editing workload by 70% using Gemini + FFmpeg on FastAPI/S3.
  • Shipped Whisker, an LLM pet-care assistant in 2 weeks with a containerized RAG backend (Postgres + FastAPI) and sub‑second response streaming; scaled to 50+ weekly beta users.
  • Designed context-aware recommendations combining planning context with chat history, increasing acceptance of suggested edits by 25%.
FastAPI FFmpeg Postgres LLM

Machine Learning Researcher

TReNDS • Atlanta, GA

May 2024 - Aug 2025
  • Built 3D ML pipelines for cognition prediction on a 12k+ subject MRI dataset, turning neuroimaging into predictive insights.
  • Designed a custom 3D CNN with Multi‑Head Self‑Attention + Squeeze‑and‑Excitation, achieving 0.34 correlation (+15% vs. CNN baselines).
  • Optimized distributed training with PyTorch DDP on a 4× GPU HPC cluster (SLURM), achieving 7× faster epochs with robust logging/checkpointing.
  • Published at IEEE ISBI 2025 and EMBS 2025, strengthening explainability of cerebellum‑related cognition models.
PyTorch DDP SLURM Neuroimaging

Computer Vision Engineer

MORSE Studio • Atlanta, GA

Aug 2023 - Apr 2024
  • Designed a mmWave radar perception pipeline using Range/Doppler FFT analysis to characterize material signatures; released open‑source experiments.
  • Built a C‑based UDP socket tool for TI mmWave radar (IWR16843 + DCA1000) enabling faster capture and multithreaded stream handling.
  • Prototyped event‑driven imaging with Luxonis neuromorphic sensors; decoded high‑frequency LED signals using Lomb‑Scargle and Python concurrency.
mmWave C/C++ Signal Processing

C++ Engineer I

Harman International • Bangalore, India

Oct 2020 - Jun 2023
  • Delivered production‑grade C++ navigation algorithms for premium OEMs (Mercedes, BMW, Audi), impacting real‑time routing at scale.
  • Engineered optimized A* navigation (C++14), resolving 350+ critical bugs and delivering 10+ major features under Agile/Scrum.
  • Reverse‑engineered map APIs to classify ambiguous road links; improved routing accuracy in dense urban networks.
  • Implemented cyclic node elimination to reduce search frontier and ECU usage; applied RAII, multithreading, and smart pointers.
C++14 Algorithms Embedded

Computer Vision Intern

KPIT • Pune, India

May 2018 - Jul 2018
  • Built a real‑time sensor‑fusion pipeline (Raspberry Pi, camera, Arduino, CAN) to stream multi‑modal data for CV training, saving 5 hrs/week.
  • Developed an offline tracking tool using YOLOv2 + KCF/Kalman, eliminating manual labeling and saving 2 hrs per dataset.
OpenCV YOLOv2 Python

Featured Work

A selection of projects that showcase my technical expertise

Cre8able — Multimodal Video Editing

Built and launched from idea to live beta in 6 weeks. Cut creator editing workload by 70% with Gemini + FFmpeg on FastAPI/S3. Context-aware recommendations improved edit acceptance by 25%.

Gemini FFmpeg FastAPI S3

Whisker — LLM Pet-Care Assistant

Shipped in 2 weeks. Containerized RAG backend (Postgres + FastAPI) with sub-second response streaming. Scaled to 50+ weekly active beta users.

RAG LLM FastAPI Postgres
Try Live Demo →

Reddit MCP Server

Model Context Protocol server for Reddit integration. Enables AI assistants to interact with Reddit's API for fetching posts, comments, and user data in a structured way.

MCP Reddit API Python AI Tools
View on GitHub →

Multi-Agent Debate for Movie Consensus

Multi-agent framework with LLM agents engaging in adversarial dialogue to refine reasoning. LLM-as-Judge evaluation improved coherence over single-agent baselines. RAG with ChromaDB + Gemini 1.5 Flash for personalized recommendations.

LLM RAG ChromaDB Gemini
View on GitHub →

Visual Py-SLAM Toolkit

End-to-end modular SLAM pipeline from raw video to trajectory visualization. Implemented ORB feature detection/matching, essential matrix decomposition, RANSAC outlier rejection, and scale initialization for robust camera trajectory estimation.

SLAM Computer Vision Python ORB
View on GitHub →

Statistical ML for Fronto-Cerebellar Circuitry

Preprocessed sMRI/dMRI from ABCD dataset (N=10K) using SPM12, DARTEL, FSL. Built Bayesian Ridge, SVR, Neural Networks for cognitive prediction. Achieved r² of 0.065 ± 0.009 with nested 5x5 cross-validation, outperforming fronto-parietal models.

Neuroimaging ML Bayesian SPM12
View on GitHub →

Detection of Milli-sized Objects with mmWave Radar

Detected millimeter-sized objects using 60 GHz IWR1683ISK FMCW mmWave radar. Detected penny-sized objects at 2m distance. Analyzed Range/Doppler FFTs across 6 materials and 5 distances using MATLAB Radar Toolbox and Python.

mmWave Radar MATLAB Signal Processing
View on GitHub →

Pose-Assisted TrackFormer

Enhanced TrackFormer (TUM/Facebook) by integrating pose estimation for improved identity consistency. Developed Pose Encoding Module with Keypoint R-CNN extracting (17 x 3) pose features, fused with track embeddings via learnable linear layer.

Object Tracking Transformer PyTorch MOT
View on GitHub →

Articles & Research

Technical articles, blog posts, and research contributions

Fronto‑Thalamo‑Cerebellar Circuitry in Predicting Cognition and Behavior of ABCD Adolescents

2025

IEEE ISBI 2025 (1‑page abstract), EMBS 2025 (4‑page paper)

Validated the cerebellum’s role in cognition and advanced explainability for neuroimaging‑based prediction models.

View

Let's Connect

Feel free to reach out through any of these channels

Phone
+1 (404) 203-1809
LinkedIn
linkedin.com/in/pvdsan
GitHub
github.com/pvdsan
Email
pvdsan@gmail.com