May 2025 – Present · 1 yr · Delhi, India · On-site

• AI for Social Good

Jan 2024 – Sep 2024 · 8 mos · Montreal, Quebec, Canada · On-site

• Worked with the Aftermarket wing to forecast parts demand and improve inventory management with reinforcement learning.

Mar 2023 – Mar 2024 · 1 yr · Montreal, Quebec, Canada · Hybrid

• Supported 2023 NextAI startups in building their AI products by recommending tools, services, best practices, and potential directions for scientific exploration.
• Drafted technical roadmaps for startups, enhancing project clarity and execution
• Conducted interviews for CDL Montreal, ensuring startups aligned with the program's technical objectives.

Nov 2022 – Dec 2023 · 1 yr 1 mo · Montreal, Quebec, Canada · On-site

• Working with Prof. Aaron Courville on building algorithms that can discover non-exploitable cooperative policies in general-sum games.
• Earlier worked with the Applied ML Research Team on using Multi-agent RL models to control signalized arterial intersections. Evaluated the scope of reinforced antagonistic agents in improving the robustness of these RL agents by simulating adversarial traffic scenarios.

Apr 2022 – Aug 2022 · 4 mos · Greater Bengaluru Area (Remote)

• Worked on forecasting orders to optimize usage of Meesho’s performance marketing budget using 150+ features representing user’s browsing activity and demographic.
• Built PySpark and Dask pipelines to train our model on ~80M installs over a span of three months

Oct 2021 – Apr 2022 · 6 mos

• Worked on building AI-powered features within the Freshworks CRM to assist sales agents
• Built robust multi-tenant Airflow training and prediction pipelines to scale Deal Insights to ~3000 accounts across different regions, actively generating ~1 million predictions every hour

May 2019 – Oct 2021 · 2 yrs 5 mos

• Built & deployed Deal Insights, a feature that predicts the success of active deals with interpretable insights about their progress. Trained “lite” models with pooled data to enable insights for customers with as little as 10 closed deals and customisable account-specific models for larger accounts.
• Built and deployed Deal Sentiment, a feature that predicts deal closure sentiment based on emails exchanged between agents and their customers during a sales funnel. Emails, despite being the most common form of communication in B2C deals, are untapped by most deal scoring services.
• Devised a 3-layer hierarchical embedding approach to scale our NLP models to accounts from diverse domains and regions. This also helped solve the cold-start problem, enabling new accounts with mere 200 mails and above to use our feature.
• Ideated and working on building a module to enrich leads using email signatures at an internal hackathon.
• Built a candidate-scoring model to assist HR in reaching out to the best applicants faster

Dec 2018 – Apr 2019 · 4 mos

• Benchmarked various lead enrichment sources, which would feed into our Sales AI pipelines.
• Worked on building an MVP for an unsupervised similar deal recommender to assist agents.

Jun 2018 – Jul 2018 · 1 mo · Bengaluru, Karnataka, India

• Worked on compression of Neural Networks with Bayesian Inference (Variational Dropout)
• Implemented alternative Linear and Convolution Layers in PyTorch to learn posterior weights using
• sparsity-inducing priors. Used these to train a YOLO v2 with Variational Dropout.
• Compressed and ran a LeNet architecture on an FPGA.
• Achieved 609x compression (Pruning + Quantization) of LeNet on MNIST. (32x more than the paper
• we implemented) and experimented with compression of VGG and AlexNet.

May 2018 – Aug 2018 · 3 mos

• Developed an application (invoked as a MIME handler) with a GUI that can place a phone call, display details about a number, save it to an address book and trigger an intent on an Android app (a fork of kdeconnect) that could be used to call. 50k+ Downloads from PyPI (as of May. 2021).

Sep 2017 – Feb 2018 · 5 mos · Bengaluru, Karnataka, India

• Implemented a Bilateral Multi-Perspective Matching model to predict similarity between Questions,
• to improve performance of FAQ chatbots.
• Worked on predicting relevancy scores between Question-Answer pairs.
• Built a crawler to fetch FAQs from a website given CSS Selectors of a QA pair.

Jun 2017 – Jul 2017 · 1 mo · Chennai, Tamil Nadu, India

• - Worked on estimating the orientation (roll-pitch-yaw) of an object given an arbitrary image (RGB, without markers)

Jan 2017 – Jun 2017 · 5 mos · Chennai Area, India

• Developed a chatbot to answer queries and handle bookings over the GoBumpr in-app chat.
• Trained an ensemble model on a limited FAQ list to determine confidence levels of different replies.
• Created a framework of JSON templates to display a variety of message bubbles, cards and buttons inspired by Facebook's Messenger API.
• Established chat-based workflows to funnel service queries into bookings.
• Libraries used : spaCy, Keras & sklearn, Pandas, Django (backend hosted on an AWS instance)