Sumedh Kothari

Machine Learning Intern

• Designed a novel adversarial ML framework combining GANs and Reinforcement Learning to autonomously generate harmful prompts and discover jailbreak vulnerabilities in modern LLMs (Claude Haiku, Qwen), with ongoing work toward an arXiv publication
• Engineered an end-to-end agentic research pipeline in Python (PyTorch, vLLM, HuggingFace) that autonomously processed 1,000+ arXiv papers and synthesized emerging red-teaming methodologies into novel attack strategies
• Built a scalable evaluation framework generating 3,000+ adversarial prompts with benign twin datasets, benchmarking 8 state-of-the-art open-source and proprietary LLMs into automated customer-facing safety reports

Research Intern

• Applied Bayesian statistical modeling using Python and NumPy/Pandas to optimize lignin valorization reaction conditions
• Utilized NMR, SEM, and HPLC modeling data to characterize value-added chemicals from lignin breakdown
• Completed literature reviews on aromatic compound synthesis, 3D printing custom GO structures, and optimal catalyst design while collaborating on manuscript for publication

Embedded Systems Engineer

• Designed and implemented a real-time lap tracking module in modern C++ for Georgia Tech’s Formula SAE Racing team, leveraging Protocol Buffers (Protobuf) for low-latency telemetry serialization and deterministic inter-process communication across vehicle subsystems
• Developing and deploying a YOLO-based object detection pipeline for driverless perception, enabling real-time cone detection and localization through C++ inference integration into the autonomous control stack

Team Lead — Mechanical

• Led a 5-member mechanical team overseeing end-to-end design and CAD of a 1 ft³ DAC system, resulting in a fully integrated prototype
• Designed and fabricated DAC hardware using 3D printing and CNC machining to build airflow pathways and zeolite containment
• Engineered the DAC system for a 75% ambient CO₂ reduction by optimizing contactor geometry and airflow–sorbent interaction, prepared for AIChE national conference evaluation

Researcher

• Designed and tested fluorinated Mg₂(dobpdc) metal-organic frameworks for vehicle exhaust CO₂ capture, improving capture potential by 20% and stability by 115%
• Characterized adsorption efficiency and regeneration performance using TGA, NMR, and cyclic adsorption experiments
• Presented findings at SYNOPSYS 2025

Simulation Intern

• Used MATLAB and dimensionless differential equations to model species concentrations in eutrophic freshwater ecosystems
• Employed sustainability over sets calculation to determine equilibria and stability of environment
• Co-authored research paper on environmental systems modeling under UCLA faculty mentorship

PolyUMA — Semiconducting Polymer Property Predictor

Graph neural network materials-property model (inspired by Meta's eSEN) trained on 6M+ OPoly26 datapoints to learn structure–property relationships in semiconducting polymers, force fields, and interatomic potentials. Built a physics-informed molecular representation pipeline (backbone planarity, π–π stacking, chain rigidity) and an iterative self-refinement training loop with domain-seeded anchoring. Submitted to the Adaptation Autoscientist Challenge.

Python · PyTorch · Graph Neural Networks · CUDA

SportSplat — 4D Gaussian Splatting for Sports

GitHub ↗

4D Gaussian Splatting volumetric reconstruction pipeline that transforms multi-view third-person sports footage into player-specific first-person renderings. Applies 3D Gaussians onto SMPL-X human frames to recreate NFL footage from any player perspective, with a differentiable rendering pipeline and camera-pose optimization (PyTorch/CUDA) for real-time synthesis.

Python · PyTorch · CUDA · Gaussian Splatting · SMPL-X

GameSense - Winners - TechHack Valley 2026 Hackathon - Hardware Track

GitHub ↗ · Live ↗

Engineered a real-time accessibility wearable for deaf gamers converting spatial audio → multi-channel haptic feedback via directional vibration. Built full hardware stack (ESP32, 9-axis IMU, 8-motor array, 3D-printed enclosure) and wrote C++ firmware for PWM motor control and UART communication. Implemented audio amplitude → haptic intensity mapping and Kalman-filtered head-tilt → WASD motion input.

Python · C++ · Embedded Systems · Fusion · WASAPI

IEEE Robotech 2026 Hackathon — Autonomous Track

GitHub ↗ · Live ↗

Designed and CAD-modeled a fully autonomous debris-clearing robot in SolidWorks. Developed a YOLO-based CV model in PyTorch achieving 95% debris detection accuracy. Integrated perception, navigation, and actuation systems for end-to-end autonomous operation.

Python · PyTorch · YOLO · SolidWorks · Embedded Systems

AiApply — AI Job Applier

GitHub ↗ · Live ↗

Full-stack AI-powered internship application platform automating end-to-end job search workflows — resume optimization, cover letter generation, programmatic Handshake submission, recruiter scraping, and automated personalized outreach at scale.

Python · Selenium WebDriver · HTML · JavaScript

Sports Betting Arbitrage Engine

GitHub ↗ · Live ↗

Aggregates live odds from multiple sportsbooks via API, normalizes data, and identifies risk-free arbitrage opportunities across H2H, Spread, and Total markets with up to 50% daily ROI. Production-ready pipeline with Dockerization and real-time analysis.

Python · REST APIs · Docker

EcoFlow

GitHub ↗ · Live ↗

Full cradle-to-grave life cycle analyzer for Amazon products. Includes Sankey diagram visualization and scalable backend architecture.

Python · FastAPI · SQLite · Upstash Redis · Next.js

GitHub Analyzer

GitHub ↗ · Live ↗

Analyzes user commit history for code quality, consistency, and hireability metrics.

Python · Next.js

ResumeForge

GitHub ↗ · Live ↗

AI-assisted resume tailoring system that adapts resumes to specific job descriptions using parsing and structured transformation.

Python · Next.js

Sustainable Farming Crop Growth Predictor

Implemented AI/ML models (MLP, linear regression, KNN, decision trees) to analyze crop growth from 30+ bacterial strain features. Optimized predictive pipeline to identify high-performing farms by soil type, achieving 160% higher accuracy over random baseline.

Python · NumPy · Pandas · Scikit-learn

ML-Assisted Particle Swarm Optimization · Completed

PACESHIP Scholar Research Grant

Awarded a $1,000 grant to investigate ML methods for black-box optimization. Built a hybrid framework integrating Gaussian Process surrogate modeling, LightGBM-guided particle repositioning, and adaptive exploration–exploitation to steer stagnating particles. An autoresearch loop autonomously ran experiments and refined strategies, achieving an average 28.8% improvement over baseline PSO across benchmarks (97.8% Sphere, 99.9% Rastrigin, 90.4% Ackley, 58.8% Rosenbrock).

PyTorch · Gaussian Processes · LightGBM · Surrogate Modeling · Optimization

Bayesian Optimization for Lignin Valorization · Completed

Georgia Tech Renewable Bioproducts Institute — Tong Laboratory

Applied Bayesian statistical modeling to optimize lignin valorization reaction conditions. Integrated NMR, SEM, and HPLC modeling data for structural characterization of value-added chemicals from lignin breakdown. Collaborating on manuscript for peer-reviewed publication.

Bayesian Modeling · NMR · SEM · HPLC · Python · NumPy · Pandas

Fluorinated MOFs for CO₂ Capture · Completed

Dressen Laboratory

Designed and tested fluorinated Mg₂(dobpdc) metal-organic frameworks for vehicle exhaust CO₂ capture. Improved capture potential by 20% and stability by 115%. Characterized performance using TGA, NMR, and cyclic adsorption experiments. Presented at SYNOPSYS 2025.

MOF Synthesis · Fluorination · TGA · NMR · Materials Science

Short-Chain PFAS Remediation Technologies · Completed

Independent Research

Conducted 35-page literature review assessing chemical and economic feasibility of PFAS removal techniques for environmental remediation. Collaborated with Yale's Dr. Julianne Rolf to validate research findings. Presented at the Symposium of Rising Scholars.

Environmental Chemistry · Literature Review · Remediation

GreenGrade — Sustainability Interval Index ↗ · Completed

Gemini API Developer Competition

Directed 20-person UCLA team to architect a fuzzy logic-driven Sustainability Interval Index calculator with scalable data integration pipelines. Developed and deployed GreenGrade (Flutter/C++), automating sustainability analytics for 175 mining firms through Gemini API.

Flutter · C++ · Gemini API · Fuzzy Logic · Data Pipelines

Freshwater Ecosystem Eutrophication Modeling · Completed

Hydrogen Engineering Research Consortium

Used MATLAB and dimensionless differential equations to model species concentrations in eutrophic freshwater ecosystems. Employed sustainability over sets calculation to determine equilibria and stability. Co-authored research paper under UCLA faculty mentorship.

MATLAB · Differential Equations · Ecosystem Modeling · Stability Analysis