# Avaya Aggarwal

> ML Engineer / Researcher building reliable multimodal systems and research tools: multimodal generation workflows, evaluation tooling, computer-vision research, and backend infrastructure. Open to ML engineering, computer vision, and research-facing roles.

- Site: https://www.avayaaggarwal.com
- Email: aggarwal.avaya27@gmail.com
- GitHub: https://github.com/OnePunchMonk
- LinkedIn: https://www.linkedin.com/in/avaya-aggarwal/
- Kaggle: https://www.kaggle.com/avayaaggarwal27
- Education: B.Tech in Computer Science Engineering with AI, Netaji Subhas University of Technology (2022 - 2026)

## Industry Experience

- ML Engineer Intern, Eternal (Dec 2025 - Jun 2026, Gurugram): AI Gateway for language, image, and video models; multimodal marketing-content workflow (planning, avatar, voice, image, video, post-production); async media processing, retries, monitoring, cost tracking. Python, LLM APIs, SQS, DynamoDB, FFmpeg.
- Software Engineering Intern, Wells Fargo (Jun 2025 - Aug 2025, Hyderabad): format-agnostic regression suite, configurable hierarchical JSON mapper with 40+ unit tests, GenAI query-and-summary workflow.
- Research Intern, Northeastern University (Oct 2024 - Mar 2025, remote): fine-tuned CLIP for melanoma detection, 56% to 84% accuracy. https://github.com/OnePunchMonk/CLIP-for-Melanoma-Detection
- Research Intern, Queen Mary University of London (Jul 2024 - Sep 2024, remote): multimodal violence detection in video, frame-level annotation.
- Software Engineering Intern, Ciena (May 2024 - Jul 2024): Python libraries for 10G PON testing, REST API integration.

## Projects

- [erasus](https://pypi.org/project/erasus/): machine unlearning framework with 41 strategies, 24 coreset selectors, benchmarks, certification, 465 passing tests.
- [AgentQuant](https://github.com/OnePunchMonk/AgentQuant): failure-aware research harness for quant strategy experiments. Typed ReAct loop, SQLite experiment memory. LangGraph.
- [TradeRecon](https://github.com/OnePunchMonk/TradeRecon): Kafka-driven real-time post-trade reconciliation with Prometheus/Grafana.
- [KoRA](https://github.com/OnePunchMonk/KoRA): multiplicative inter-adapter communication for PEFT, with formal analysis and tests.
- [uois_toolkit](https://pypi.org/project/uois-toolkit/): unified PyTorch toolkit for five unseen-object instance-segmentation datasets, with IRVL, UT Dallas.
- [tiny-tribe](https://github.com/OnePunchMonk/tiny-tribe): distilled ~14M param multimodal brain encoding model predicting fMRI responses to video (Pearson r 0.7278).
- [Pix Coreset](https://github.com/OnePunchMonk/PixCoreset): coreset generation for CV, 90% data compression for KNN segmentation. FastAPI, Streamlit, Docker.
- [Forget Me Net](https://github.com/OnePunchMonk/ForgetMeNet): verifiable SISA unlearning with 121s shard-level unlearning and proof certificates.
- [Coresets for Image Segmentation](https://github.com/OnePunchMonk/Coresets-for-Image-Segmentation): hackathon project.
- [BTCTrade](https://github.com/OnePunchMonk/BTCTrade): LSTM + technical-indicator trading bot, backtested.

## Current work

- [Diffusion post-training lab](https://github.com/OnePunchMonk/diffusion-post-training-lab): reproducible LoRA, preference optimization, GRPO, and step distillation on open-weight diffusion models. Looking for feedback on evaluation design, reward quality, and ablations.

## Skills

- Languages: Python, SQL
- ML and data: PyTorch, TensorFlow, benchmarking and evaluation
- Backend and infra: FastAPI, PostgreSQL, Docker, Kafka, REST APIs
- Engineering practice: Pytest, regression testing, research prototyping, system design