Dec 2023 – Jan 2025 · 1 yr 1 mo

• Worked on building and productionising realtime ML solutions for e-commerce Ad Fraud Detection for billion-scale data on Sponsored Ads and Twitch.
• Developed algorithms for fraud detection such as (a) GPT-style model which takes in user-level activity sequences, (b) R-GCN model to detect colluding entities committing fraud and (c) DNN-based algorithms which use activity-based features to catch click and impression bots.

Sep 2022 – Dec 2023 · 1 yr 3 mos

• Built and deployed a DNN-based algorithm, which uses activity-based features (click frequency and velocity) and distils model scores into interpretable fraud entity lists to detect sophisticated click bots.
• Collaborated on building a realtime hybrid system of heuristics and ML algorithms to detect sophisticated crawler bots using network request level features, leading to detection of crawler impressions. Paper accepted at KDD Workshop on AI-Enabled Cybersecurity Analytics.

Jul 2022 – Sep 2022 · 2 mos

• Built a networking performance testing library from scratch to compare different networking stacks.
• Used by entire team to compare multiple stacks like Open vSwitch (OVS) and OVS-Data Plane Dev Kit (OVS-DPDK), and make important decisions like which stack to migrate to, etc.

Aug 2021 – Jul 2022 · 11 mos

• Implemented an end-to-end pipeline for fetching counters for Access Control Lists (ACLs) for telemetry purposes.
• Designed the APIs, RPCs, and implemented the logic for fetching the counters by interfacing with the OVS and the Open Virtual Network (OVN).

May 2022 – Sep 2022 · 4 mos

• Worked with the Keras team [Google] on KerasNLP, Keras’ native NLP offering and a multi-backend Keras, i.e., new Keras version able to run on top of PyTorch, TensorFlow, JAX.
• Top contributor to KerasNLP, involved at every stage of development, and made important contributions – models like DistilBERT, BART, DeBERTa, Whisper, etc., efficient XLA-compatible text generation APIs, NLP metrics, LoRA and so on.
• Invited to give technical talks on Keras (and Deep Learning, in general) at Google events around the country and outside. Received the Google AI/ML Community Award and the Google Open Source Peer Bonus Award.
• Paper accepted at JMLR 2024 (Open Source track).

Mar 2022 – Nov 2023 · 1 yr 8 mos · Remote

• Worked with the Keras team [Google] on KerasNLP, Keras’ native NLP offering and a multi-backend Keras, i.e., new Keras version able to run on top of PyTorch, TensorFlow, JAX.
• Top contributor to KerasNLP, involved at every stage of development, and made important contributions – models like DistilBERT, BART, DeBERTa, Whisper, etc., efficient XLA-compatible text generation APIs, NLP metrics, LoRA and so on.
• Invited to give technical talks on Keras (and Deep Learning, in general) at Google events around the country and outside. Received the Google AI/ML Community Award and the Google Open Source Peer Bonus Award.
• Paper accepted at JMLR 2024 (Open Source track).

Jul 2021 – Sep 2022 · 1 yr 2 mos · Remote · Remote

• Built an ICD coding framework which provides a streamlined pipeline for preprocessing, training, and evaluation for the automatic ICD coding task, including an interactive demo and explanation of outputs using popular interpretability methods.
• Paper accepted at EMNLP 2022.

Mar 2021 – May 2022 · 1 yr 2 mos · Goa, India · Remote

• Worked on infusing super-pixel-based knowledge (higher-order perceptual groups of pixels) into deep neural networks for the image classification task.
• Used a CNN to deal with spatial information present in the image, and a Graph Attention Network (GAT) model to deal with relational superpixel information.
• Won the best short paper award at ACM-SE 2022.

Jan 2021 – Sep 2021 · 8 mos · Delhi, India · Remote

• First exposure to Vision+Language models.
• Focused on building a dataset of political memes - built a local annotation tool, and curated memes data from Reddit.
• Designed multimodal models (ViT + BERT, VLBERT, etc. as encoders) to locate political explanations of a given meme in a piece of text.

Apr 2020 – Nov 2020 · 7 mos

• Developed a novel architecture dubbed "3DBLES‑UNet" which uses the concept of "Bilateral Grids" for estimation of depth maps with prominent, accentuated edges.
• Paper accepted at IC3D 2020.

Jan 2020 – May 2020 · 4 mos

Aug 2019 – Dec 2019 · 4 mos

May 2019 – Jul 2019 · 2 mos · Pilani, Rajasthan · On-site

• Worked on a popular Deep Learning problem "Anomaly Detection in Videos of Daily Living using Deep Learning" and simulated state-of-the-art results.
• Trained a GAN on dynamic images generated from videos from the URFD dataset (for fall detection).