Sree Dhyuti Nimmagadda
AI Researcher, Clariti
I am an AI researcher at Clariti, where I work on LLM and vision-based understanding of construction documents. I hold a Master's in Artificial Intelligence from Northwestern University (GPA 3.98/4.00) and an integrated (M.Tech + B.Tech) dual degree in Computer Science from IIITDM Kancheepuram, Chennai.
My research interests span large language models, reinforcement learning, and computer vision. Recently I have worked on Image segmentation, long-context LLM evaluation (at Writer), 3D medical image segmentation (at the Abazeed Lab), multi-agent RL fine-tuning frameworks, and self-supervised world models. Outside research, I sing and produce music.
News
- 2026 Won the AWS Multimodal AI Hackathon for Code-ICU, an autonomous ML-training monitor with AI voice calls and self-healing fixes.
- 2026 1st place, Finance Track, UC Berkeley Quantum AI Hackathon, for Qbitrade.
- 2026 2nd place at the Physical AI Hackathon (makermod, San Francisco) for an ACT-based bimanual manipulation system.
- 2026 Finalist (top 8 of 70 teams) at Weave Hacks 4, Weights & Biases, for AutoRL.
- Jan 2026 Joined Clariti as an AI Researcher.
- Dec 2025 Completed my MS in Artificial Intelligence at Northwestern University (GPA 3.98/4.00).
- Jun 2025 Started as an AI Research Intern at Writer, San Francisco, working on long-context evaluation.
- Dec 2024 Presented my paper on LLM evaluation metrics at IEEE TENCON 2024, Singapore.
Publications
-
Improved Text-Summarization with PEGASUS and Siamese Network Evaluation
IEEE TENCON 2024 (IEEE Region 10 Conference), Singapore
Selected Projects
Code-ICU [code]
Autonomous ML-training monitor that detects gradient explosions and loss anomalies with an LLM agent pipeline, phones you via an AI voice call, and applies self-healing code fixes with corrected re-runs.
Winner, AWS Multimodal AI Hackathon.
Reinforcement Learning & Embodied AI
AutoRL [code]
Multi-agent RL fine-tuning framework that converts natural-language tasks into automated RL experiments across PPO, SAC, A2C, and GRPO, with self-healing training infrastructure and W&B Weave + Redis observability.
Weave Hacks 4 finalist.
Multi-Task Embodied Learning for Bimanual Robot Manipulation
Full-stack embodied AI system built on LeRobot — 78 teleoperated demonstrations (~51K frames), ACT and SmolVLA policies for multi-task behavior cloning across utensil stacking and liquid pouring.
2nd place, Physical AI Hackathon.
LLMs, NLP & AI Agents
REINFORCE-based Contradiction Correction in LLMs [code]
BERT-based contradiction detector fine-tuned with LoRA and improved via a REINFORCE RL framework — ~90% detection accuracy, reducing contradiction rate from 65% to 44%.
AI-Powered Hotel Recommendation [code]
GPT-4o with agentic retrieval and RAG to reduce hallucinations, delivering real-time, preference-aligned hotel recommendations.
QnA Bot [code]
TF-IDF based parser that reads paragraphs and answers questions about them.
Computer Vision
CrowdPhysics [code]
Crowd-crush early warning from any camera: optical flow into a self-supervised world model that flags danger as "surprise," with Claude-agent risk reasoning, model-based RL, and pre-event 3D crowd-flow simulation. Latent probe recovers boundary stress at R² 0.94.
Diffusion Modeling from Scratch [code]
Conditional Denoising Diffusion Probabilistic Model (DDPM) implemented from scratch for image generation.
Machine Learning & Data Science
Qbitrade [code]
Quantum-enhanced trading pipeline: 14 alpha models, a hybrid QSVM meta-labeling framework, and a PPO portfolio optimizer improving Sharpe from 0.859 to 1.038 (+21%). 1st place, UC Berkeley Quantum AI Hackathon.
Experience
-
Jan 2026 – Present
AI Researcher, Clariti Remote, USA
LLM-based floorplan understanding: raised plan-tagging recall from 68.6% to 82.6% with Vertex AI foundation models; built a multi-stage YOLO segmentation pipeline (>90% precision); halved model calls per prediction. -
Jun – Sept 2025
AI Research Intern, Writer San Francisco, CA
Long-context evaluation framework spanning 18 datasets over 7 tasks; LLM-as-a-judge and few-shot evaluation validated against human comparisons; attention-mechanism interpretability. -
Mar – Jun 2025
AI Researcher, The Abazeed Lab Chicago, IL
DynUNet-based 3D segmentation of 117 organs-at-risk (94% cross-validation Dice), deployed in the lab's radiation-oncology workflow; benchmarked Swin UNETR and ViT. -
May – Oct 2023
ML Research Intern, BioSystems & Controls Lab Chennai, India
Semi-supervised autoencoder regression for real-time fermentation monitoring (validation R² 0.61 → 0.89); NIR-spectrum regression for apple sugar prediction. -
Aug – Dec 2022
Machine Learning Intern, Tiny Banyan Technologies Chennai, India
YOLOv5 for real-time pothole and crack detection (99% accuracy); trained an intern team on GCP workflows. -
May – Jul 2021
Theoretical Research Intern, IIITDM Kancheepuram Chennai, India
Novel NP-complete algorithms for Steiner tree problems in Split, Interval, and Chordal graphs (~2× faster runtime).
Education
-
2024 – 2025
MS, Artificial Intelligence - Northwestern University, Evanston, IL. GPA 3.98/4.00.
NLP, deep learning, GenAI, graph neural networks, data science, high performance computing. -
2019 – 2024
M.Tech + B.Tech, Computer Science - IIITDM Kancheepuram, Chennai, India. CGPA 8.86/10.00.
Thesis: "Improved Text-Summarization with PEGASUS and Siamese Network Evaluation" (IEEE TENCON 2024).
Awards & Honors
- 2026 Winner, AWS Multimodal AI Hackathon (Code-ICU)
- 2026 1st Place, Finance Track, UC Berkeley Quantum AI Hackathon (Qbitrade)
- 2026 2nd Place, Physical AI Hackathon, makermod, San Francisco
- 2026 Finalist (top 8 of 70 teams), Weave Hacks 4, Weights & Biases (AutoRL)
- 2022 Global Finalist (top 45 of 1,500 teams), OpenCV AI Competition
Writing
-
Medium
Digital Image Processing: A Primer (Part 1) — fundamentals of digital image processing for computer vision.
-
Medium
Neural Networks That Learn Without Activation Functions — architectures that learn without traditional activations.
Music
Outside research, I sing and produce music on FL Studio. A few tracks from my SoundCloud:
Contact
The fastest way to reach me is email: sreedhyutin@gmail.com. I am always open to conversations about research collaborations, ML engineering roles, or interesting problems in AI. You can also find me on LinkedIn and GitHub.