Feb 2021 – Jun 2022 · 1 yr 4 mos

• Mission: drive down abuse driven by fake, compromised and inauthentic accounts across Facebook, Messenger and Instagram. Team's scope involves classifying billions of data points every day to protect user activity across posting, commenting, messaging etc by leveraging a wide variety of heuristics, ML and clustering techniques. Managing a team of 9 engineers working on backend/ML

Jul 2018 – Feb 2021 · 2 yrs 7 mos

• We aim to prevent people from abusing the platform to misrepresent themselves, artificially increase viewership/popularity or distribute content en masse.

May 2017 – Jul 2017 · 2 mos · Menlo Park, California

• Worked with the Media Search team to improve the experience/time spent.

Aug 2015 – Dec 2015 · 4 mos · Singapore

• Thesis in Machine Learning. Researched and developed new techniques for multi-label classification.

May 2015 – Jul 2015 · 2 mos · Karlsruhe Area, Germany

• Architectular space exploration of superscalar processors, advised by Prof. Muhammad Shafique.
• . Extended the out-of-order superscalar CPU model (for fault-tolerance) in gem5, an open-source simulator for Computer Architecture research
• Provided protection against soft errors using DMR with rollback and TMR (Triple/Double Modular Redundancy) for various pipeline components
• TMR and DMR can be selectively enabled for different pipeline components based on their spatial and temporal vulnerability to soft errors while executing several benchmark applications
• Compared the default gem5 checkpointing with DMTCP (Distributed Multithreaded Checkpointing) based checkpointing, coupled with different checkpoint compression mechanisms, for the periodic checkpoints made for DMR. Comparison parameters include checkpoint sizes, checkpoint saving and recovery time
• Analyzed the area and power overhead due to the above heterogenous hardening modes, using data obtained from McPAT, to enable soft error mitigation under area/power/performance constraints