Atharva Walawalkar

Experience

Chewy Inc., Bellevue, WA — Applied AI Intern (Jun 2025 – Aug 2025)

Architected M.A.T.T. and P.A.W.S., a user intent analysis and memory system on a Neo4j graph transforming 4.5B+ clickstream events into persistent agent memories for behavioral analysis across internal teams and customer applications.
Built Ad-visor, an autonomous diagnostic agent using a Reason-Critique-Analyze loop that evaluates ad campaign performance per user and auto-escalates patterns to marketing teams via Slack integration.
Engineered a Fluid Frontend Engine to dynamically rerank homepage widgets based on navigation patterns, achieving sub-50ms personalization latency with LLM-in-the-loop architecture.
Presented multi-agent portfolio to 50+ stakeholders including CEO, CTO, and VP of AI — earned Runner-up & "Most Innovative Use of AI" awards.

Hackathons & Competitions

Google Gemini Hackathon — Commentator: Real-Time AI Game Caster [Dec 2025] — 3rd Place, $20,000

Autonomous agent providing live play-by-play commentary for esports broadcasts with no human intervention. Reduced broadcast latency from 15s to 7s with entropy-based frame filtering via FFmpeg. Chained Gemini 2.5 Flash Lite for high-frequency vision analysis with ElevenLabs TTS.

OpenEnv / PyTorch Hackathon — WolfeClick: RL Agent for Competitive Pokemon [Mar 2026]

Built an OpenEnv-compatible gym wrapping Pokemon Showdown via WebSocket with structured markdown state representation and JSON action validation. Trained a Qwen3-4B LoRA adapter using GRPO on live battle trajectories with a 7-component dense reward signal including KO incentives, type-effectiveness bonuses, and a -10.0 hallucination penalty. Designed an SFT → rollout → GRPO pipeline with measurable win-rate improvement; deployed battle replay viewer on HF Spaces.
github | hf space | model | blog

AWS Builder Hackathon — AI Interview Tutor [2025]

Voice-first technical interview simulator where the candidate speaks and draws on a whiteboard; a Claude Sonnet critique agent and Claude Haiku emotion router run in parallel, triggered after a 1.5s speech pause. Streaming transcription via AWS Transcribe WebSocket with Minimax TTS for emotionally-matched feedback. Dual-agent async design on AWS Bedrock.

Llama Hackathon — Cerebral Valley × Meta [2025]

WikiFace: open-source platform inspired by DeepWiki, tailored for Hugging Face. Uses Llama's extended context window for fine-grained model information (papers, architecture trees, tokenization notes) with a Llama-powered chatbot interface.

Cal Hacks 11.0 [36 Hours, 2025]

EasyQuizzes: RAG-powered study assistant with LLaMA 3.2-11B-vision OCR and ChromaDB vector embeddings for question generation.
github

COEP Neurohacks [48 Hours, 2023]

English-to-Hinglish translator using a PyTorch-based NLP pipeline. Lightweight Gradio UI and efficient CPU inference.

Projects

Edge-AI Flood Alert System [Provisional Patent Filed, 2024]

Custom LSTM trained on 40 years of weather data, quantized via TFLite for fully offline inference on Raspberry Pi. 0.73mm RMSE. Offline-first architecture serving predictions over local WLAN during flood-induced internet blackouts. Led a 3-person team in collaboration with the Indian Meteorological Department for live flood monitoring in Mumbai suburbs.

Ongoing Projects

Adaptive RL for Dynamic ROI Selection in rPPG (M.S. Thesis) 2025–2026

Clean-room PPO rebuild for semantic ROI12 facial region selection with a MediaPipe-first preprocessing stack.
State: semantic ROI12 schema (64 columns). Action: continuous (12,) scores → top-k weighted ROI mixture (k=3 default).
Reward modes: hr_mae_sparse and rbs_calibrated (sparse only).
Canonical pipeline: preprocess videos → semantic ROI12 state CSVs → generate aligned GT → train PPO → evaluate with local windowed or paper-like 20s metrics.
Training on MMPD, V4V, PURE datasets; benchmarking against R2I-rPPG and SAMC baselines.

Gator Tutor — Backend Lead SFSU Software Engineering Capstone

Stood up the entire FastAPI backend from scratch — the foundation the 6-person team built on.
Designed the complete booking & scheduling system: availability slots, duration-based expiry, status filtering, and optimized DB queries with joinedload to fix N+1 problems.
Built tutor search with fuzzy matching, separate course/tutor endpoints, and comprehensive filters by subject and availability.
Set up all DevOps infrastructure: SSH tunnels for local DB access, AWS EC2 deployment, NGINX reverse proxy, Google Analytics.
Owned the media handling system for profile photos and hardened production — fixed routing bugs, role update logic, soft deletes, and chat read-flags.

github

Research

Deep Learning Flood Alert System [Patent Pending, 2024]

IoT and deep learning system combining LSTM networks with edge computing over Raspberry Pi for real-world environmental monitoring.

Credit Card Fraud Detection [Publication, 2023]

"A Comparative Analysis of Machine Learning Techniques for Credit Card Fraud Detection." Evaluated ensemble methods (XGBoost, Random Forest), traditional ML (SVM, Logistic Regression), and anomaly detection techniques.

Technical Stack

ML/AI: PyTorch, HuggingFace Transformers, scikit-learn, TFLite, ONNX, LoRA, GRPO, PPO, LangChain, DSPy

Infra & Languages: Python, C++, JavaScript, SQL, AWS EC2/Bedrock, Docker, FastAPI, Neo4j, Git, Linux

Education

M.S., Data Science & AI | San Francisco State University | Expected May 2026

GPA: 4.0/4.0
Relevant coursework: Natural Language Processing, Algorithms, Operating Systems

B.Tech., Electrical, Electronics & Communication Engineering | Pune University | 2024

GPA: 8.2/10.0
Relevant coursework: Signal Processing, Electromagnetism, Computer Networks, Digital Communication

Contact

Looking for full-time AI research, Applied AI, or MLE opportunities starting May 2026. Feel free to reach out if you want to discuss AI, collaborate on projects, or talk about opportunities.

atharvawal27@gmail.com | San Francisco, CA