Working on the LLM team at Anyscale wearing many hats.

Current project: Core member of the NovaSky project (Jan 2025 - Present).

• One of the core contributors of SkyRL: https://github.com/NovaSky-AI/SkyRL , building a full stack library for post-training LLMs.
• Integrated FlashRL with SkyRL, enabling training with quantized rollouts: https://skyrl.readthedocs.io/en/latest/examples/flash_rl.html
• Core contributor to https://novasky-ai.notion.site/skyrl-v0 - implemented a scalable remote server for RL training on SWE-Gym, contributed to building asynchronous multi-turn rollout implementation to improve SWE-Bench performance of Qwen-3-8B by 5.8%.
• Core contributor to https://novasky-ai.notion.site/skyrl-sql , one of the first open-source models trained with multi-turn RL on Text2SQL - matching GPT-4o and o4-mini on the Spider benchmark.
• One of the core maintainers for the Skythought repo: https://github.com/NovaSky-AI/SkyThought/commits?author=SumanthRH
• Led building the evaluation suite for SkyThought.

Previous project: LLMForge (May 2024 - December 2024) Worked on new fine-tuning features, performance improvements, product CLI/SDK, etc

• Added support for different fine-tuning tasks (such as instruction tuning and causal LM) training as well as function-calling fine-tuning.
• Improved model support to allow bringing any HuggingFace model with any chat template to fine-tune on Anyscale.
• Led building the LLM Models SDK for easily going from fine-tuning to serving on the platform.

Worked on the Generative AI team at C3!

• Set up a finetuning codebase from scratch for C3’s Generative Search application
• Features: Support for difference causal and sequence-2-sequence models, ability to mix different training datasets (for a text-to-text or a causal language modelling task), visualize metrics on multiple evaluation datasets, etc
• Trained 10B+ parameter models on 1M+ samples using DeepSpeed and 🤗 Accelerate.

This was a part time internship I did with Hakimo in Spring'23. This was my first time working on video-based models, so that was fun!

• Worked on video-based object detection models for Hakimo’s Remote Guarding Solution
• Trained 3D ResNets on Hakimo’s video surviellance data and experimented with single and muli-pathway SlowFast networks.

Served as a Teaching Assistant for CSE 232: Principles of Database Systems and CSE 21: Mathematics for Algorithms and Systems. Was a lot of fun, resposibilities included:

• Conducting weekly discussion sessions for 50+ students.
• Preparing question papers for midterm and final examinations.

Bachelor’s Thesis.

• Demonstrated fast reconstruction of a 12 frame video from a single image of a lensless camera, reducing inference time from 2 hours to 30 milliseconds.
• Proposed an efficient reconstruction framework - a physics-aware neural net
  trained in an adversarial fashion, used feature-based loss for photorealism.
• Employed a trainable inversion layer to reverse the forward process of the camera, along with a UNet for perceptual enhancement.

• Implemented a face detection algorithm for KYC services.
• Trained a Multi-task Cascaded Convolutional Neural Network using > 200,000 images.
• Reduced false positives 10 times and false negatives by 2.5 times.
• Employed hard positive mining, data augmentation to reduce recall by 5%.

Presented on Ray and it’s applications as scalable agent infrastructure at the Agentic AI Summit in UC Berkeley

Watch on YouTube →

Guest lecture for CSE 234 (Graduate Computer Systems) covering advanced topics in distributed systems and machine learning infrastructure.

Watch on YouTube →

• Specialization in AI/ML
• JUMP mentor for underclassmen at UC San Diego.
• MicroMBA from the Rady School of Management.

• My bachelor’s thesis was on video reconstruction for lensless cameras.
• President’s Gold Medal
• Computer Vision and Intelligence Club