Applied AI
At OLX, I work on machine-learning systems for marketplace quality and discovery—from framing the problem to evaluating models, running experiments, and deciding what should ship.
Read the work01 / Core work
A marketplace has to make millions of small judgments: Is this listing safe? Is it genuine? Is it good enough to show? What needs a human review?
My work sits between those product questions and the models used to answer them. That means understanding policy and operational constraints, choosing useful signals, evaluating failure modes, and helping turn model output into a decision the product can safely use.
I worked on systems that combine text and image signals to assess marketplace listings. The work included image understanding, NSFW detection, quality signals, and decisions about when a model should act automatically or support moderation.
I built and evaluated ways to find repeated and near-duplicate listings across text and images. This included similarity methods and image hashing, with the product goal of reducing repeated content without blocking legitimate ads.
Trust and safety is not one model. I have worked across user, ad, and chat moderation, where policy, adversarial behavior, review capacity, precision, recall, and latency all pull in different directions.
02 / Core work
A recommendation system can improve its headline metric while making the marketplace feel narrower. I care about both the immediate result and the distribution it creates.
I have worked on recommendation experiences across homepages and ad-detail pages in classifieds and motors. The work involves model comparison, experiment design, result analysis, and questions that are harder to compress into one metric—particularly diversity and the balance between relevance and discovery.
Homepage and ad-detail recommendation surfaces
General classifieds and motors marketplaces
Offline evaluation and controlled product experiments
Relevance, discovery, freshness, and diversity trade-offs
One example
I used KL divergence to compare the distributions produced by recommendation approaches. It gave us a way to discuss diversity as an observable model property rather than a subjective impression of the feed.
Earlier work
Before OLX, I worked in PwC’s innovation team from 2019 to 2021. My projects included pose detection, call analysis, and automating parts of PowerPoint production. It was where I learned to move between a messy business request and a technical system someone could actually use.
Background
2021—present
OLX
Applied machine learning for marketplace quality and discovery
2019—2021
PwC · Innovation team
Analytics, computer vision, and automation
Education
PGDBA
Post Graduate Diploma in Business Analytics from IIM Calcutta, IIT Kharagpur, and ISI.