╔══════════════════════════════════════════════════════════════════╗
║ ║
║ ████████╗██╗ ██╗███████╗██╗ ██╗ █████╗ ██████╗ ║
║ ██╔══╝██║ ██║██╔════╝██║ ██║██╔══██╗██╔══██╗ ║
║ ██║ ██║ ██║███████╗███████║███████║██████╔╝ ║
║ ██║ ██║ ██║╚════██║██╔══██║██╔══██║██╔══██╗ ║
║ ██║ ╚██████╔╝███████║██║ ██║██║ ██║██║ ██║ ║
║ ╚═╝ ╚═════╝ ╚══════╝╚═╝ ╚═╝╚═╝ ╚═╝╚═╝ ╚═╝ ║
║ ║
║ T Y A G I · ML Engineer · Researcher ║
╚══════════════════════════════════════════════════════════════════╝
tushar = {
"role" : "ML Engineer · Researcher · Builder",
"education" : ["M.S. CS @ Georgia Tech (4.0 GPA)", "B.Tech CS @ IIT Mandi"],
"focus" : ["Agentic AI", "Computer Vision", "Multimodal LLMs", "RL"],
"previously" : "ML Engineer II @ Adobe (4 years)",
"patent" : "US20240404070A1 — Deep ML for book page boundary detection",
"location" : "Atlanta, GA 🍑",
"currently" : "Building things that push the frontier 🚀"
}| 🎯 Metric | 📊 Impact |
|---|---|
| 🚀 Feature MAU Boost | +50% via real-time CV classification pipeline |
| 👥 Users Revitalized | 3M+ dormant users re-engaged |
| 📱 Monthly Active Users | 5M using Book Mode feature I spearheaded |
| ⚡ Inference Speedup | 3× faster · 4× smaller via quantization & pruning |
| 🔥 Query Accuracy | +40% with Chain-of-Thought prompting in Acrobat AI |
| 📈 Downsampling Speed | 10× faster than Lanczos — custom C++ algorithm |
🧠 AI / ML Core
⚙️ Systems & MLOps
🔬 Research Domains
Computer Vision · Agentic AI · Multimodal RAG · RL / GRPO · Autonomous Driving
Model Optimization · QLoRA / PEFT · Self-Supervised Learning · OCR / VLMs · Sensor Fusion
🎬 Agentic Video Retrieval with Structured Verification — click to expand
Compute-optimal multimodal retrieval resolving the ingest-vs-query cost bottleneck.
Applied adaptive candidate pruning + uncertainty-aware tool selection to maximize information gain per GPU dollar.
LangGraph·Multimodal RAG·Pydantic·Agentic AI
🧩 VLM Context Engineering for Video Analytics
Task-type classifiers that dynamically route across 5 RAG-to-agentic retrieval strategies.
+10% accuracy and 2× speedup over static methods on Video-MME benchmark.
LangGraph·VLMs·Multimodal RAG
📜 Historical Manuscript OCR — Open Source Contribution
Benchmarked 8 Vision-Language Models (Qwen 2B–32B) for handwritten text recognition.
Achieved 0.124 CER with Qwen2.5-VL-7B · 52% CER improvement via LoRA fine-tuning.
PyTorch·LoRA·Transformers·Quantization
🚗 DriveContrast — Robust Autonomous Driving
Fine-tuned VLM-based autonomous driving pipeline (AutoVLA + VideoMAE) on Waymo Open Dataset.
+10.8% trajectory accuracy on hard ADAS scenarios under sensor perturbations.
VLM·QLoRA·PyTorch·HPC/SLURM·Spatiotemporal Encoding
🎯 LLM Fine-Tuning with GRPO
Implemented Group Relative Policy Optimization for LLM alignment.
Used group-mean rewards as control variates + KL-divergence penalties for stable policy updates.
Reinforcement Learning·Policy Optimization
"I don't just train models — I build systems that scale,
optimize for what actually matters, and ship things
that real people use."
— From 3M revived users to a US Patent,
from IIT Mandi to Georgia Tech Llamas Lab.
The loop never closes.
╔═══════════════════════════════════════════╗
║ Always building. Always optimizing. ║
║ Open to research collabs & cool ideas 🤝 ║
╚═══════════════════════════════════════════╝