# Avaya Aggarwal: full profile

> I turn ambiguous ML ideas into testable systems: multimodal generation workflows, evaluation tooling, computer-vision research, and backend infrastructure. I care about the evidence behind a result as much as the implementation.

Short index: https://www.avayaaggarwal.com/llms.txt

## Working principles

- **Build for inspection.** Typed interfaces, tests, traces, and useful failure modes—not just a polished demo.
- **Evaluate before claiming.** I prefer reproducible baselines, explicit assumptions, and honest boundaries around results.
- **Bridge research and systems.** From model experiments to async services, observability, and workflows people can actually use.

## Experience

### ML Engineer Intern, Eternal (Dec 2025 - Jun 2026, Gurugram, India)

Built applied-AI and backend infrastructure for multimodal generation and AI platform operations.

- Worked on the AI Gateway that connected the platform to language, image, and video models through a common interface for routing, validation, and model configuration.
- Built a multimodal marketing-content workflow that took a campaign brief through planning and review to avatar, voice, image, video, and post-production outputs.
- Added the supporting backend for asynchronous media processing, retries, persistent state, monitoring, and cost tracking.
- Worked on computer-vision and image-generation workflows for food, catalogue, and brand-personalisation use cases.
- Built Streamlit demos for creating and editing marketing content and reviewing costs.

Stack: Python, LLM APIs, SQS, DynamoDB, FFmpeg

### Software Engineering Intern, Wells Fargo (Jun 2025 - Aug 2025, Hyderabad, India)

Built internal tooling around regression validation, mapping workflows, and GenAI-assisted analysis.

- Reduced QA effort with a format-agnostic regression suite.
- Designed a configurable hierarchical JSON mapper backed by 40+ unit tests.
- Implemented a query-and-summary workflow for business users to speed up analysis.

Stack: Python, Test Automation

### Software Engineering Intern, Ciena (May 2024 - Jul 2024, Gurugram, India)

Improved internal testing workflows for network systems using Python and APIs.

- Revamped in-house Python libraries for 10G PON testing.
- Integrated REST APIs to improve maintainability and execution efficiency.

Stack: Python, Test Automation

## Projects

### AgentQuant

- Category: Agentic AI
- Status: open source, active
- Link: https://github.com/OnePunchMonk/AgentQuant
- Page anchor: https://www.avayaaggarwal.com/#project-agentquant
- Stack: Python, LangGraph, SQLite

**What it is.** A failure-aware research harness for proposing, testing, and retaining lessons from quantitative-strategy experiments.

**Result.** Built around a typed ReAct loop, SQLite-backed experiment memory, reproducible runs, and an inspectable offline demo.

### TradeRecon

- Category: Trade reconciliation engine
- Status: open source
- Link: https://github.com/OnePunchMonk/TradeRecon
- Page anchor: https://www.avayaaggarwal.com/#project-traderecon
- Stack: Python, Kafka, SQLite, Prometheus

**What it is.** A Kafka-driven post-trade reconciliation engine that compares execution, broker confirmation, and PnL records in real time.

**Result.** Flags price, quantity, timestamp, and PnL mismatches with SQLite audit trails, Prometheus metrics, Grafana dashboards, and HTML reports.

### KoRA

- Category: PEFT
- Status: research prototype
- Link: https://github.com/OnePunchMonk/KoRA
- Page anchor: https://www.avayaaggarwal.com/#project-kora
- Stack: PyTorch, PEFT, Vision Models

**What it is.** A research exploration of multiplicative inter-adapter communication for parameter-efficient fine-tuning.

**Result.** Includes an explicit formal analysis, counterexamples, tests, and an implementation boundary between the proposal and validated behaviour.

### uois_toolkit

- Category: Open-source computer vision toolkit
- Status: shipped (PyPI)
- Link: https://pypi.org/project/uois-toolkit/
- Page anchor: https://www.avayaaggarwal.com/#project-uois-toolkit
- Stack: Python, PyTorch Lightning
- Collaborators: Jishnu Jayakumar (https://jishnujayakumar.github.io/), Animesh Maheshwari (https://animeshmaheshwari22.github.io/)

**What it is.** A three-person collaboration with researchers at IRVL, UT Dallas: a unified PyTorch toolkit for unseen-object instance-segmentation datasets, loaders, sampling, augmentation, and evaluation.

**Result.** Standardizes five UOIS datasets behind one API, includes built-in research metrics, and shipped as an installable PyPI package.

### tiny-tribe

- Category: Brain encodings
- Status: proof of concept
- Link: https://github.com/OnePunchMonk/tiny-tribe
- Page anchor: https://www.avayaaggarwal.com/#project-tiny-tribe
- Stack: Python, PyTorch, Distillation, fMRI

**What it is.** A lightweight distilled brain encoding model that predicts fMRI responses to naturalistic video from audio, visual, and text features.

**Result.** Distills TRIBE v2 into a compact multimodal student model with about 14M trainable parameters and reported Pearson r of 0.7278 on the proof of concept.

### erasus

- Category: Machine unlearning framework
- Status: shipped (PyPI)
- Link: https://pypi.org/project/erasus/
- Page anchor: https://www.avayaaggarwal.com/#project-erasus
- Stack: Python

**What it is.** A machine unlearning framework built around coreset selection, multiple unlearning strategies, evaluation tooling, and certification workflows.

**Result.** Includes 41 unlearning strategies, 24 selectors, benchmark pipelines, documentation, and a test suite with 465 passing tests.

### Pix Coreset

- Category: Coreset proof of concept
- Status: proof of concept
- Link: https://github.com/OnePunchMonk/PixCoreset
- Page anchor: https://www.avayaaggarwal.com/#project-pix-coreset
- Stack: Python, Computer Vision, Docker, FastAPI, Pytest, Streamlit

**What it is.** An interactive coreset generation tool for computer vision experiments, packaged with a FastAPI backend and Streamlit interface.

**Result.** Achieved 90% data compression for KNN segmentation workflows with Dockerized deployment, CI, and Pytest coverage.

### Forget Me Net

- Category: Unlearning proof of concept
- Status: proof of concept
- Link: https://github.com/OnePunchMonk/ForgetMeNet
- Page anchor: https://www.avayaaggarwal.com/#project-forget-me-net
- Stack: Python, FastAPI, Streamlit, Transformers, IPFS

**What it is.** A verifiable machine unlearning prototype built around SISA training, model serving, and downloadable proof artifacts.

**Result.** Reached 121-second shard-level unlearning and produced downloadable proof certificates backed by FastAPI, Streamlit, Transformers, and IPFS.

### Coresets for Image Segmentation

- Category: Hackathon project
- Status: hackathon
- Link: https://github.com/OnePunchMonk/Coresets-for-Image-Segmentation
- Page anchor: https://www.avayaaggarwal.com/#project-coresets-for-image-segmentation
- Stack: Python, Machine Learning, Computer Vision, Segmentation

**What it is.** A coreset-driven image segmentation project focused on reducing training data while preserving segmentation quality.

**Result.** Built as a focused project submission around coresets for image segmentation.

### BTCTrade

- Category: Hackathon project
- Status: hackathon
- Link: https://github.com/OnePunchMonk/BTCTrade
- Page anchor: https://www.avayaaggarwal.com/#project-btctrade
- Stack: Python, LSTM, Backtesting, Technical Indicators

**What it is.** A deep-learning trading bot that combines LSTM price forecasting with technical indicators including Bollinger Bands and moving averages.

**Result.** Validated with backtesting, reporting RMSE of 417, Sharpe ratio of 8.71, and Sortino ratio of 2.47.

## Current work

### Diffusion post-training and evaluation

Building reproducible workflows for LoRA, preference optimization, GRPO, and step distillation on open-weight diffusion models.

Looking for: Feedback on evaluation design, reward quality, ablations, and reproducible baselines.
Link: https://github.com/OnePunchMonk/diffusion-post-training-lab

## Education

B.Tech in Computer Science Engineering with AI, Netaji Subhas University of Technology, 2022 - 2026.
Coursework: Data Structures and Algorithms, Machine Learning, Artificial Intelligence, Deep Learning, Operating Systems, Database Management Systems, Game Theory.

## 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

## Contact

- Email: aggarwal.avaya27@gmail.com
- GitHub: https://github.com/OnePunchMonk
- LinkedIn: https://www.linkedin.com/in/avaya-aggarwal/
- Kaggle: https://www.kaggle.com/avayaaggarwal27