Nov 2024 – May 2025 · 6 mos

• Architected and shipped an AI-driven employee-recognition service (TypeScript + React + PostgresQL) to web, iOS, Android, MS Teams, and Zoom in 3 months; drove full product roadmap and release. Launched various recognition types - peer kudos, birthdays, and anniversaries boosting engagement. Integrated Okta SCIM and gifting platform Gifted, enabling one-click SSO login and instant gift-card payouts.

Nov 2023 – Oct 2024 · 11 mos

• 1. Managed numerous tracks in the layout service (LS), which dictated the layout of pages (trays, titles, ads, UI, etc) in 26 countries.
• 2. Revamped LS to introduce reusability concepts in widgets and spaces, which removed the need to create every page from scratch. This brought > 50% reduction in page creation time.
• 3. Used Retool to evolve the internal LS dashboard, doubling the efficiency, reducing errors by 5x, and allowing seamless migration from the old dashboard to the new one.

May 2022 – Oct 2023 · 1 yr 5 mos

• 1. Made an open-source contribution to Apollo GraphQL Router by adding Automated Persisted Queries, which saves bandwidth by reducing HTTP request payload sizes in orders
• of magnitude.
• 2. Decoupled all teams’ deployments by a major refactor, which vastly improved the speed of CI/CD in their services.
• 3. Developed HotBites as a hackathon project, enabling users to make short clips (in Instagram Reel/TikTok format) of content on Hotstar and share with friends to improve engagement.
• 4. Enabled the Because you watched feature by introducing the concept of placeholder widgets, which are trays that get their data in real-time.

May 2021 – Jul 2021 · 2 mos

• Made an instrumentation library for the Go HTTP client that accurately keeps track of the client's requests and connections.
• 1. Kept account of the number of idle connections, active connections, waiting requests, as well as request wait times.
• 2. Extended previously available option of tracking a single request using HTTPTrace by providing statistics on client level.
• 3. Exposed previously hidden, low-level errors for better debuggability if the response returns an error on a higher level.

Apr 2020 – Jun 2020 · 2 mos · Gurgaon, Haryana, India

• Worked on making machine learning models more explainable and interpretable by giving reasons for decision making and defining input-output relations to follow. Broadly worked on two topics:
• 1. Monotonic Neural Networks- This involved enforcing monotonicity of the output with a given subset of input features. Applied Deep Lattice Network, Unconstrained Monotonic Networks, and a point-wise Loss that penalizes non-monotonic trends.
• 2. Explainable AI- Implemented a multitude of methods to give reasons for specific predictions of models and understanding of the whole decision-making process in general. Otherwise called local and global methods. Also, applied multiple model-specific and model-agnostic methods to explain decisions involving structured, image and text data. Some of the methods applied are LIME, SHAP (Shapely Values), and Class Activation Maps (CAM).