Abhijay Salvi

Stores Foundational AI

• Building RL as a Sandbox service to optimize GPU allocation and save compute costs during LLM training.
• Developing CLI tool integration with Claude Code and Codex to decrease idea-to-deployment times.

Reinforcement Learning · LLM Training · GPU Scheduling · CLI Tooling

Machine Intelligence

• Developed scripts for anomaly detection and statistical analysis for guidance systems using PyTorch, NumPy, and Pandas.
• Led codebase restructuring for a 7-person agile team; managed sprint planning and tracking in Jira.
• Applied an internal LLM tool to automate System-on-a-Chip and hardware-in-the-loop testing by 50%.
• Tracked prompt-response pairs, error rates, and usage to debug failure patterns and improve internal LLM reliability by 15%.

PyTorch · NumPy · Pandas · LLM Tooling · Jira

Resilience Intelligence

• Built and deployed via GCP a 3-agent LangGraph pipeline for supply chain facility discovery, reducing research time by 70%.
• Engineered a PostgreSQL facility database and cache indexing of 100K+ records across 100+ countries.

LangGraph · GCP · PostgreSQL · Multi-Agent Systems · Python

Physical Automation & Robotics

• Implemented boundary detection and real-time control for autonomous warehouse navigation using the Boston Dynamics SDK.
• Built a React/Next.js + REST API dashboard to monitor and control Boston Dynamics Spot warehouse operations.
• Trained a PyTorch + OpenCV model for real-time package overflow detection, achieving 95% precision.
• Presented technical solutions to C-suite executives.

Boston Dynamics SDK · PyTorch · OpenCV · React · Next.js · REST APIs

Waresport

• Directed a global cross-functional team through iterative product development, scaling from 0 to 100+ users.
• Led design and implementation of LLM and AI-based scheduling, cutting management time by 90%.
• Architected full-stack systems and delivered a custom CMS for dynamic content publishing.
• Integrated CI/CD pipelines to streamline deployment.

React · Node.js · TypeScript · LLM Integration · CI/CD · AWS

Innovation and R&D

• Built a Flask research platform for 20+ Marketing and PR members, reducing data aggregation workflows by 75%.
• Integrated a CustomGPT chatbot across internal and public web apps, improving retrieval efficiency by 85%.
• Re-architected a Python + PyTorch NLP pipeline for machinery tagging, boosting throughput by 150%.

Flask · Python · PyTorch · NLP · Node.js · FastAPI

AI Operating System for Sports Clubs

• Built a TypeScript + Python + MongoDB SaaS platform with a 7-person team, scaling to 20+ clients.
• Designed a rule-based scheduling algorithm with OpenAI API integration, reducing management time by 90%.
• Deployed to production on AWS with GitHub Actions CI/CD and Docker containers.

TypeScript · Python · MongoDB · OpenAI API · AWS · Docker · GitHub Actions

Hacklytics — 2nd Place, Sports Track

• Built a custom RAG pipeline with the Gemini LLM to generate contextual AI coaching insights from real-time fight analytics.
• Engineered a low-latency asyncio pipeline processing live video via OpenCV + PyTorch at 95% accuracy.
• Surfaced injury prevention insights and concussion risk metrics from biomechanics, hardware-free.

Python · OpenCV · MediaPipe · PyTorch · Gemini API · Django · Taipy

Hazard Alert & Visual Emergency Network

• Built an AI hazard intelligence platform using Next.js + Flask, converting crowdsourced photos into verified real-time incidents.
• Integrated the Google Maps API for dynamic safe-zone routing, computing nearest shelter and evacuation path using A*.
• Used multi-source verification so a single bad report can't create a false incident.

Next.js · Flask · Roboflow · Google Maps API · SQLite · Python

PennApps — 1st Place AR/VR · 3rd Overall

• Achieved 97% tumor detection accuracy by fine-tuning a pretrained model for real-time MRI frame analysis with AR overlay.
• Architected a full-stack Firebase web app with an integrated 3D brain tumor modeling system in Matplotlib for exploration.

AR/VR · Computer Vision · Firebase · Python · Matplotlib · iOS

Hackathon

• Built AR experiences for memory preservation and reconnection.
• Combined SnapAR and Unity with a Python backend.

SnapAR · Unity · Python · AR/VR

Hackathon

• Built stories that evolve from user choices using generative AI.
• Used Google Vertex AI and NLP to keep branching narratives coherent.

GenAI · Google Vertex AI · NLP · Web

Hackathon

• Designed a secure digital identity flow with privacy-first defaults.
• Integrated Azure and Firebase for authentication and storage.

React · Azure · Node.js · Firebase

NextTech CSforGood — Statewide Winner

• Used OpenCV to detect open parking spots from lot imagery.
• Built a mobile app showing live availability, cutting time spent circling for a space.

OpenCV · Python · Computer Vision · Mobile

Research Team Member — Georgia Tech Europe

• Improved transcription clarity by 20% via an ASR and wav2vec2 pipeline that detects and replaces unclear phonemes.
• Focused on preserving meaning while raising intelligibility for hearing aid users.

Deep Learning · wav2vec2 · ASR · PyTorch · Python · Speech Processing

Research Team Member — Medical Robotics @ Georgia Tech

• Applied responsible AI principles to user safety within real-time feedback control loops.
• Built real-time Python ML models for EMG classification, improving prosthetic responsiveness.
• Worked across embedded control systems and simulations in an agile research team.

Python · Machine Learning · Embedded Systems · Real-time Control · Simulation

Team Member — Big Data Big Impact @ Georgia Tech

• Built a large-scale pipeline for crop forecasting using satellite time-series and ML classification.
• Improved yield prediction accuracy by 18% through tuning and feature optimization.

Python · Machine Learning · Satellite Data · Time-Series · Feature Engineering

Researcher under a CalTech Ph.D. Candidate — Independent Research

• Developed and benchmarked state-of-the-art models on NOAA datasets, improving drought prediction accuracy by 18%.
• Compared random forest, nonlinear regression, SVM, KNN, and neural networks across local and global scales.

Python · Scikit-Learn · Neural Networks · Pandas · NumPy

Research Participant — UC Santa Barbara Summer Research Academy

• Designed a novel soft robot to cleanse coral reefs of microplastics.
• Programmed Series Pneumatic Artificial Muscles (SPAMs) with Arduino C.
• Presented the work at a UCSB research seminar.

Arduino · C · Soft Robotics · Pneumatics