Aditya Lal

I’m a Software Engineer Intern focused on LLM training and reinforcement learning (RLHF), with a strong bias toward production reliability. At Media.net, I automated a continual training workflow that schedules training and evaluation, validates results, pushes artifacts to Google Cloud Storage, and restarts GPU-backed vLLM deployments on a tight cadence.

I also build internal tooling that helps teams move faster. I created a centralized experiment-tracking backend API with a Streamlit dashboard for training and performance telemetry, reducing manual reporting effort from 5 days to 1 day (80%) for cross-team work.

Quality and correctness are central to how I ship. I wrote pytest integration tests across 9 config-loading scenarios, reaching 95% code coverage and catching parameter and file-path misconfigurations before training runs.

On the modeling side, I extended production RLHF with reward modeling and GRPO policy training stages (evaluated against PPO) for ad ranking. I improved offline RPM by 15% and achieved a 3.5% online RPM lift in A/B tests, while also building multi-LLM evaluation tooling with Llama, Qwen, Gemma, and ModernBERT to raise evaluation accuracy from 50% to 65% and accelerate reward model training.