Sep 2024 – Feb 2026 · 1 yr 5 mos · Remote

• RAG Evaluation
• Working on introducing an evaluation system for RAG applications for easy debugging and performance improvement - to be integrated with Truefoundry's Cognita, an versatile open-source RAG framework.
• Agents
• Building custom agentic systems for different use cases for our clients.

Jun 2023 – Sep 2024 · 1 yr 3 mos

• Job recommendation system
• Designed the candidate generation and ranking modules for job feed recommendation system for >1 million WAU - for improving the % transacting users and the number of applications per TU.
• Recommendations also used to power communications via PNs and WhatsApp - aiding in reactivations and fulfilment.

May 2022 – Jun 2023 · 1 yr 1 mo

• Candidate Trust and Safety
• Built a Data Science model to flag fraudulent employers on the platform to ensure safety for our candidates.
• This, along with some proactive fraud markers are being used to prioritise investigations.
• Impacted into a reduction of ~25% of complaint volume on the platform.

Jan 2021 – Apr 2022 · 1 yr 3 mos

• Truck Utilisation
• Used Computer Vision techniques on the live video of the truck loading process to determine effective truck utilisation - with an accuracy of about 80%.
• Developed an instance segmentation model to identify object masks in individual frames and performed post-processing on them.
• Customer Intent Categorisation
• Automated grievance handling by implementing a customer intent categorization model on emails with 80% accuracy.
• The model is being used to classify the intent for all concerns Delhivery gets through email in order to ensure seamless customer experience.

Nov 2017 – Jan 2021 · 3 yrs 2 mos

• Damage prediction modeling
• Implemented a binary classification model to model the probability of damage of a shipment given attributes like shipment contents, the source and destination of the shipment.
• The model had a recall of 56% on flagging top 5% most fragile shipments. Flagged shipments are marked and carefully handled by the operations team.
• Product Categorisation:
• Implemented text categorisation model to compute the main/sub category in a hierarchical tree structure, given product name.
• Model is used to find the categories for ~ 1 million/day Delhivery shipments with ~ 95% accuracy.
• PIN code polygons:
• Used k nearest neighbour classifier to remove noisy points from GPS data for delivery field executives.
• Constructed disjoint polygons for Indian postal codes by creating concave hull around the clean points.
• The ~ 15000 polygons are now used for cleaning GPS data and various last mile optimizations.
• Return prediction:
• Created predictor variables and built a random forest model for predicting the return probability of each shipment.

Jul 2015 – Sep 2016 · 1 yr 2 mos · Gurgaon, Haryana, India

• Recommendation Framework for a major Indian Bank:
• Created predictor variables using historic spending behaviour and asset ownership data. Recommended financial products to customers using collaborative filtering technique.

May 2014 – Jul 2014 · 2 mos · London, England, United Kingdom

• Investigated the pre-melting behavior of metals using molecular dynamics simulations in DL POLY CLASSIC.

May 2013 – Jul 2013 · 2 mos · Hong Kong SAR

• Investigated reaction of polymerization of Br-functionalised Porphyrins using STM and analyzed images using Excel and Origin Pro 8.
• Simulated the system for various temperatures using FORTRAN codes.