## Shreyas Basavatia — AI/ML Engineer & Researcher

> Source: shreyasportfolio.co

## Overview

Shreyas Basavatia is a Computer Science student at Georgia Tech specializing in integrating machine learning and AI into real-world applications, with focus areas in multimodal AI, natural language processing (NLP), and reinforcement learning (RL). Experience spans AI research (IBM Research; Georgia Tech's EI and HCAI Research Lab), production engineering (Founding Engineer at Everyday Media), and enterprise software (Software Engineering Intern at Venerable). Relevant dual-use domains include emergency response, fire detection/public safety, real-time video/audio analytics, and intelligent document processing.

## Skills and Capabilities

- **Applied ML/AI:** Multimodal AI, NLP, and reinforcement learning across real-world applications.
- **LLM systems:** LLM-agnostic Multi-Agent AI Service supporting custom agent creation and tool use (Everyday); multi-stage LLM pipelines; RAG and Chain-of-Thought prompting; agentic RAG.
- **Model fine-tuning:** Fine-tuned open-source multimodal models from the Qwen, Gemma, and LLaMA families; integrated agentic RAG for dermatological diagnosis.
- **Computer vision:** Fine-tuned YOLOv5 for fire/person detection; Google Cloud Vision image labeling.
- **Real-time processing:** Low-latency video/audio capture systems analyzing hundreds of hours of data in real-time (Stealth Startup); Socket.IO real-time video feed processing (Flame AI); real-time semantic search.
- **Speech/audio:** OpenAI Whisper speech-to-text handling background noise and varying accents; Twilio 911 call audio capture; automatic meeting transcription.
- **Big data pipelines:** AWS Kinesis, AWS Glue, Parquet, Iceberg, and AWS Athena for real-time user-behavior analytics.
- **Full-stack development:** Next.js, React.js, TailwindCSS, shadcn/ui, Flask, Django, Hono.js; databases PostgreSQL, MySQL, MongoDB (Prisma, Drizzle ORMs); Clerk authentication; Stripe payments; Firebase; Supabase.
- **Cloud & AI services:** AWS S3, Lambda, Bedrock, Textract, Comprehend; Anthropic Claude 3.5 Sonnet; Google Gemini 1.5 Flash; LangChain, HuggingFace, Gradio, PyTorch, Scikit-learn, Pandas, NumPy.
- **Embedded/IoT:** Arduino + Raspberry Pi + Force Sensitive Resistor sensing systems (TEDuP).
- **Game development:** Unreal Engine (C++), Unity (C#), level design, and blueprints.

## Projects

### Flame AI — Fire Detection & Public Safety System
Advanced fire detection system that alerts first responders and admins quickly for optimal resource allocation and swift rescues. Uses a fine-tuned YOLOv5 model for visual fire/person detection and a logistic regression model for non-visual detection (temperature, humidity). On detection, it automatically alerts authorities with fire location and number of people affected. Real-time processing via Socket.IO; built with Flask, Scikit-Learn, YOLOv5, HTML/CSS/JavaScript. Serves first responders, residential areas, educational institutions, and commercial spaces. (Hacklytics 2024)

### QuickER — Emergency Response & ER Management
End-to-end solution improving emergency response from the initial 911 call through ER treatment. Performs live 911 call analysis (Twilio + OpenAI Whisper), predicts conditions and recommends treatments/medications/diagnostics (LangChain + Llama3.1), pre-allocates resources (medications, X-ray/MRI rooms, operating rooms), routes patients to beds, and provides a hospital-state dashboard. Frontend: Next.js (SSR, ISR, hydration) + TailwindCSS; backend: Flask. (HackGT 11)

### IntelliDoc — Intelligent Document Processing
End-to-end application parsing company term sheets, performing complex cross-referenced lookups, reasoning about scattered information, running business-date calculations, and generating insights. Supports PDF upload or selection from S3, with a choice between AWS Bedrock and open-source HuggingFace models. Built with AWS (S3, Lambda, Bedrock, Textract, Comprehend), Anthropic Claude, LangChain, Gradio, Pandas, NumPy. Built for Venerable.

### Everyday Media — Digital Asset Lifecycle Management
Production-ready platform managing 100,000+ creative assets for enterprise clients. Features AI-powered semantic search (Ocular AI) ingesting/tagging/indexing 1,000+ video/image assets daily with natural-language queries, an LLM-agnostic Multi-Agent AI Service, big-data analytics pipelines, and predictive context models across 20+ applications.

### Additional Applications
- **GraphIQ:** AI-powered interactive learning platform (data structures/algorithms) with multimodal voice/text/visual interaction, real-time Mermaid.js diagrams, infinite practice problems, exportable study guides. Uses Claude 3.5 Sonnet, Gemini 1.5 Flash, Whisper, LangChain, NLX, Next.js/React.
- **Poker Banter:** Poker bot fine-tuned to play like professional players and impersonates fun characters like Lebron James to trash talk you and your moves.
- **FotoMind:** Google Photos-like store/search app using Google Cloud Vision, Next.js/React.js, Firebase.
- **Finance Tracker:** SaaS spending-analytics platform (Next.js, React.js, Hono.js, Clerk, Drizzle, PostgreSQL, TailwindCSS).
- **Atlanta Food Finder:** Restaurant discovery app (Django, Google Cloud API, TailwindCSS).
- **Wrapify:** Spotify listening-insights app (Django, Spotify API, TailwindCSS).
- **Ecommerce Website:** Full-stack store + admin dashboard (Next.js, React.js, Prisma, MySQL, Stripe, Clerk).
- **Netflix Clone:** Full-stack streaming app (MongoDB, Prisma, Next.js, React.js).
- **Landmarks & Scrumdinger:** iOS apps in Swift/SwiftUI (Scrumdinger includes automatic meeting transcription).
- **TEDuP (Teaching Electronic Drum Pad):** Sensor-feedback hardware system (Force Sensitive Resistors → Arduino → Raspberry Pi).

## Research & Publications

- **Cultivating Multimodal Intelligence: Interpretive Reasoning and Agentic RAG Approaches to Dermatological Diagnosis** — Working Notes of CLEF 2025 (Thakrar, Basavatia, Daftardar).
- **STORY2GAME: Generating (Almost) Everything in an Interactive Fiction Game** — preprint, IEEE COG 2025 (Zhou, Basavatia, Siam, Chen, Riedl).
- **STARLING: Self-supervised Training of Text-based RL Agent with Large Language Models** — ACL Findings 2024 (Basavatia, Murugesan, Ratnakar).
- **ComplexWorld: A LLM-based Interactive Fiction Learning Environment for Text-based RL Agents** — KBCG @ IJCAI 2023 (Basavatia, Ratnakar, Murugesan).

## Performance

- **IntelliDoc:** 96% accuracy verifying term sheet information; 10x faster than the prior manual process; over 90% unit-test code coverage.
- **Everyday Media:** Semantic search reduced time to find assets by 90%; predictive models sped up creative operations by up to 50%; 1,000+ assets indexed daily; 100,000+ assets managed.
- **Georgia Tech EI/HCAI Lab:** Multi-stage LLM pipeline achieving >90% success rate compiling player-interactable actions into coherent game worlds.
- **IBM Research:** Generated 100+ training environments via STARLING; pre-trained RL agents improving performance by up to 50%.
- **ImageCLEF MEDIQA-MAGIC 2025:** Team achieved 2nd place (submission scored 6th); challenge co-organized by Microsoft, Stanford University, and Hospital Clinic of Barcelona.

## Differentiators

- **On-premise / privacy-preserving inference:** QuickER runs its entire prediction model locally so patient symptoms and call transcriptions are not sent to third parties.
- **Model-agnostic architecture:** IntelliDoc supports switching between AWS Bedrock and open-source HuggingFace models; Everyday's Multi-Agent service is LLM-agnostic.
- **Proven enterprise scale:** Solutions serving Uber, Nike, and Walmart (via Everyday Media) and financial services (Venerable).
- **Dual-use safety focus:** Fire detection (Flame AI) and emergency-response optimization (QuickER) built on citable domain research, including the Journal of Emergency Medicine and ACEP on ER crowding.
- **Full-stack + research depth:** Peer-reviewed publications combined with production deployment experience.

## Professional Experience

- **Software Engineering Intern, Flock Safety** (San Francisco, May-Aug 2026).
- **Founding Engineer, Stealth Startup** (Remote, from Jan 2026) — low-latency real-time video/audio analytics.
- **Founding Engineer, Everyday Media** (Atlanta, GA; Dec 2024–Dec 2025).
- **AI Researcher, EI and HCAI Research Lab, Georgia Tech** (Atlanta, GA; Jan 2024–Dec 2025).
- **Software Engineering Intern, Venerable** (West Chester, PA; May–Aug 2024).
- **AI Researcher, IBM Research** (Remote; Jun 2021–Jan 2024).

## Contact

- Website: shreyasportfolio.co
- Resume: ShreyasBasavatiaResume.pdf (updated January 30, 2026)
- Affiliation: Computer Science, Georgia Tech