Atharva Chandak.

ML Engineer 2

eBay

Sept 2025 - Present

Sellers describe the same product in countless ways. My work is figuring out which catalog product a listing really is, so recommendations and the buyer experience start from the right product.

• Built multimodal models that use a listing's photos and text together to find its catalog match. Better matches fed recommendations and showed up as an eight-figure lift in GMV.
• Pushed precision up while widening coverage in several categories at once, which usually means trading one against the other.
• Fine-tuning open vision LLMs (Qwen3, Gemma 4) to read the small visual details that separate one collectible from another.
• Optimised and deployed our LLMs, which raised throughput (TPS) by 30%, improved GPU utilisation and cut long-term serving cost.
• Prototyped a shopping assistant with memory and retrieval, as a first look at what agentic help could do for buyers.

SWE

Wells Fargo

Jul 2023 - Sept 2025

Two years on the merchant lending side of a large bank, split between services that move money and the first LLM product I took all the way to production.

• Started with a prototype assistant for merchant workflows and stayed with it through hardening and launch. Loan applications finished up to 30% faster.
• Worked on credit-card services that handle over a million transactions a month for thousands of merchants.
• Wrote the POS EMI microservices that let a purchase at the counter be split into installments, plus analytics so merchants can see how their business is doing.

Research Assistant

Mila / UdeM

Jan 2023 - Jun 2023

A detector trained on new conditions tends to forget the old ones, which is a problem for a car that drives through sun, rain and night. My thesis was about stopping that.

• Cut catastrophic forgetting in object detection by about three quarters.
• Tested on-device incremental learning for robots. Transformer detectors lost around 4 mAP50 points when adapting to new domains, where CNNs lost nearly 15.
• On the CLAD-D benchmark the approach beat baselines by roughly 15%.

Research Intern

Airlab, CMU

Aug 2021 - Dec 2021

TartanAviation records real aircraft around airports with cameras, ATC radio and ADS-B. Drawing aircraft boxes from scratch at that scale is impractical, so I worked on how the labels get made.

• Wrote the vision pipeline that turned footage from about 550 sequences into over 3 million high-resolution images.
• Built the camera calibration and auto-labeling framework. Calibration ties each camera to the airfield so ADS-B positions land in the image, and those projections help check the detector's output. The result was 700k+ aircraft bounding boxes.

Summer Intern

Wells Fargo

May 2022 - Jul 2022

A summer project built around one question: can a customer ask about a home loan out loud and get a useful answer?

• Built the whole voice loop. Speech-to-text takes the question in, an action server decides and composes the reply, and TTS speaks it back.
• Wrote the NLP pipeline: intent classification landed in the low 90s on F1 and entity extraction in the low 80s.
• Queries were resolved in up to about two thirds less time.

Visiting Research Intern

Visual Computing Group, Harvard

Jan 2022 - May 2022

• Contributed to PyTorch Connectomics, adding cellpose model support for neuron instance segmentation.
• Explored semi-supervised methods and designed an affinity learning + transformer pipeline to improve 3D segmentation accuracy.

Research Intern

AIRL, IISc

Jun 2021 - Dec 2021

• Worked on Generalized Continual Zero-Shot Learning, integrating incremental learning with ZSL for realistic DL deployment.
• Extended the setup to task-free, out-of-distribution generalization across diverse Computer Vision tasks.