• Architected a voice-first agentic AI browser capable of natural language web interaction, covering search, navigation, form automation, and multi-step task execution, built around real-time LLM and voice API integration with Gemini and Deepgram.

• Designed the backend infrastructure on AWS using EC2 and DynamoDB, with ElastiCache for low-latency caching to keep real-time LLM inference responsive under load.

• Managed infrastructure as code with Terraform and built CI/CD pipelines and observability tooling to support fast iteration cycles while maintaining production reliability.

• Designed a transformer-based navigation policy for autonomous agents in multi-target environments where standard RL approaches struggled with long-horizon planning, outperforming baseline methods on path planning tasks.

• Built a custom OpenAI Gym-compatible RL environment and trained agents using PPO, TRPO, VPG, and imitation learning, improving obstacle avoidance and task completion by 30%.

• Implemented policy gradient methods at the algorithm level, including advantage estimation, reward shaping, and trajectory sampling.

• Co-developed GRIP, a Boolean convex optimization model for gene regulatory network inference that exceeded state-of-the-art baselines by 20% on large-scale genomic datasets. Co-authored the resulting publication.

• Existing pipelines couldn't handle the data volume needed for large-scale motif discovery, so built and deployed high-throughput pipelines on IU's BigRed200 HPC cluster using SLURM, boosting throughput by 25%.

• Unified 4 inference frameworks (DirectNet, SCENIC+, Pando, CellOracle) into a single modular Python toolkit, cutting environment setup time by 40% across the team.

• Automated biological annotation pipelines via Ensembl and UCSC APIs, processing 30k+ records and improving analysis accuracy by 15%.

• Developed time-series and regression models to forecast hospital patient census and inflow patterns, translating noisy operational data into actionable resource planning signals.

• Ran large-scale training and hyperparameter experiments on IU's BigRed200 supercomputer, enabling fast iteration across model architectures and feature configurations.

• Designed and built autonomous and interactive robotics systems including multi-sensor navigation vehicles, a voice-interactive robot with sub-300ms response latency, and bots for robot soccer using Arduino, Raspberry Pi, and custom embedded firmware.

• Worked across communication protocols (UART, I2C, SPI), real-time control loops, and actuator systems.

• Ran workshops for 50+ students and supervised hackathon teams from hardware prototyping through to demo-ready systems.