NEWJan 2026 — Joined Llamas Lab @ Georgia Tech for Video Retrieval research MSCSAug 2025 — Started MS Computer Science @ Georgia Tech PATENTSept 2024 — Page boundary detection patent published WINJune 2023 — Winner: Acrobat Desktop Ideathon NEWJan 2026 — Joined Llamas Lab @ Georgia Tech for Video Retrieval research MSCSAug 2025 — Started MS Computer Science @ Georgia Tech PATENTSept 2024 — Page boundary detection patent published WINJune 2023 — Winner: Acrobat Desktop Ideathon
Open to research collaborations

Tushar
Tyagi

MS Computer Science @ Georgia Tech · Former ML Engineer @ Adobe · Building compute-efficient AI systems for video understanding and document intelligence.

Computer Vision Multimodal AI LLMs MLOps Video Retrieval Document Intelligence
About
Building at the edge of AI
4.0
GPA @ Georgia Tech
4 yrs
ML Engineering @ Adobe
5M+
monthly users of my patent feature
1
published US patent

I'm a grad student at Georgia Tech specialising in ML and Computer Vision. Previously I spent 4 years at Adobe shipping real-time document intelligence and computer vision models that now serve millions of users worldwide. I care deeply about compute efficiency — building systems that are not just accurate, but deployable and fast.

Projects
Selected work
🔬 Llamas Lab · Ongoing

Agentic Video Retrieval

Jan 2026 – Present

Compute-optimal multimodal retrieval resolving the ingest-vs-query cost bottleneck. Applies adaptive candidate pruning and uncertainty-aware tool selection to maximise information gain per GPU dollar.

LangGraph Multimodal RAG Agentic AI Pydantic
📊 CS 8803 LRV

VLM Context Engineering

Aug – Dec 2025 · +10% accuracy · 2× speedup

Fine-grained optimisation system using task-type classifiers to dynamically select among five context retrieval strategies — from baseline RAG to agentic search — on the Video-MME benchmark.

Video-MME RAG VLM
🛣️ VIP 6603 · Ongoing

DeepPIR

Aug 2025 – Present · +4% SOTA on IoP metric

Adapts Swin ViT models to self-supervised 2-channel pavement image registration with a 1×1 input adapter layer, enabling road-degradation monitoring for infrastructure agencies.

Swin ViT Self-supervised Transfer Learning
📝 CS 8903 Agentic AI

PROPEL

May – Jul 2025 · +26.7% NDCG@1

Training-free, explainable framework using empirical bias profiling and inverse propensity weighting to mitigate position bias in LLM re-ranking across real-world datasets.

LLM Re-ranking Bias Mitigation Explainable AI
Experience
Where I've worked
🔬
Research Assistant — Llamas Lab Jan 2026 – Present
Georgia Institute of Technology
Working on compute-efficient natural language video retrieval, focusing on agentic multimodal systems with adaptive resource allocation.
🏢
Machine Learning Engineer 2 Jul 2021 – Jul 2025
Adobe
Built real-time document classification models with 90% accuracy (+50% feature MAU). Shipped boundary detection and segmentation models (+30% accuracy). Created LLM-based AI assistant features. Reduced model sizes significantly improving on-device performance.
PyTorchTFLiteComputer VisionLLMs
🎓
Product Development Intern May – Jul 2020
Adobe
Developed smart image cleaning tools using image processing and deep learning. Created synthetic data generation scripts using VBA.
Education
Academic background
🐝
Georgia Institute of Technology
MS in Computer Science · Machine Learning
Aug 2024 – Present GPA 4.0 / 4.0
🏔️
IIT Mandi
B.Tech Computer Science · Minor in Management
Aug 2017 – Jun 2021 CGPA 8.39 / 10
Patent
Published IP
📄 US20240404070A1 ↗ · Published Sept 2024
Precise Book Page Boundary Detection
A machine learning system using convolutional neural networks to automatically detect and extract book page boundaries from photographs. Powers the "Book Mode" feature in Adobe Scan — users can extract left and right pages from book spreads in a single click.
📱 5M+ monthly active users
Recognition
Awards & achievements
🥉Third Place — Hacklanta.dev Startup Track
🏆Winner — Acrobat Desktop Ideathon, June 2023
🥈Runner-Up — Document Cloud India Hackweek, Dec 2023
Star of the Month — Adobe, September 2022
🎤Speaker — Adobe DC Summit 2022 (top 8/200)
🥉3rd Place — IIT Mandi Freshmen Innovation Challenge
Connect
Let's talk

Open to research collaborations, interesting problems, and good conversations.

Visitors