Feb 2024 – Aug 2025 · 1 yr 6 mos · Bengaluru, Karnataka, India · Hybrid

• ✤ Buying offer prediction and pricing recommendation for the inventories of core sellers like Unilever & L’Oréal and other non core sellers.
• ✤ Goal: Boost revenue by marketing high scored SKUs out of thousands and reduce offer to order conversion time by right pricing & gain more seller confidence.
• ✤ Created Unit Depletion Model for tracking liquidation through vendors, seller partners and marketplaces other than Pollen.
• ✤ Build multiple small models to answer how much % a product's inventory can liquidate, at what price, in what time, through which liquidation channels, should be domestic or export.
• ✤ Leveraged external import data from UNComtrade, Amazon and Lazada MY, IND, TH B2C pricing to enrich models.
• ✤ Data cleaning and sharing analytics related to sellers, inventory, offers, orders, conversion rates, recovery and sell thru rates.
• ✤ Final models used ‑ Ensemble of Catboost and TabNet, a Transformer based DNN.
• ✤ DS Metrics: Core ‑ Precision: 88%, Recall: 86%. Non‑Core ‑ Precision: 82%, Recall: 81%
• ✤ Model results tracking ‑ Wrote a script for tracking the actual DS metrics obtained on monthly basis after receiving the ground truths.
• ✤ Deployed model using AWS services (EC2, Lambda), Docker for containerization and FastAPI

Jan 2022 – Feb 2024 · 2 yrs 1 mo · Bangalore Urban, Karnataka, India · On-site

• ✤ Built 10 models for demand forecasting across hyperlocal & lifestyle BU (3 men’s + 2 women’s apparel SCs + 3 footwear SCs + 10 NOS SCs).
• ✤ Goal: Aid better warehouse planning, proactive seller communication, working capital reduction, and avoid liquidation losses.
• ✤ FSN‑MSKU Mapper integrated to make current production flow consistent with MSKU ids and their attributes for downstream consumption.
• ✤ Validated earlier claimed results post integration via backcasting ‑ symmetric and similar results indicated right integration.
• ✤ Men’s Apparel SCs ‑ 35% absolute WMAPE DP improvement + 8.8 Cr topline impact annually + 17% coverage improvement.
• ✤ Women’s Apparel SCs ‑ 33% absolute WMAPE DP improvement + 5.6 Cr topline impact annually.
• ✤ NOS SCs ‑ 24% absolute WMAPE improvement + 13 Cr topline impact annually + single model helps in easier maintainability for future.
• ✤ DP converged the forecasts of apparel and footwear models for inventory health planning and use cases like automated buying.
• ✤ Handled taxonomical changes, verticals to SCs remapping, combining vertical level models to single SC level model + result validation

Jul 2021 – Jan 2022 · 6 mos · Hyderabad, Telangana, India

• ✤ Detection of fraudulent medicine claims to maximize the total money recouped.
• ✤ Converted third-party’s detection rules from SQL pseudocode to PySpark code.
• ✤ Leveraged fake data of varied similarity index for the data drift simulation process
• needed to be shown on the monitoring dashboard.
• ✤ Revamped results by nearly 7% using hyperparameter optimization.

May 2020 – Jun 2020 · 1 mo

• ✤ Rewarded by PPO (Pre-Placement Offer)
• ✤ Successfully completed 3 tasks of COVID-19 detection project: Lung Segmentation, Infection Segmentation and COVID-19 Classification.
• ✤ Used OpenCV techniques in building an appropriate pre-processing & post-processing pipeline.
• ✤ Models built: UNet and UNet++ along with 3-fold and 4-fold cross-validations.
• ✤ Best results on Infection Segmentation -> Dice: 0.95, IOU: 0.92
• ✤ Best results on Lung Segmentation -----> Dice: 0.98, IOU: 0.97
• ✤ Best results on COVID-19 Classification -> AUCROC: 0.99, F1: 0.98

Dec 2019 – Jan 2020 · 1 mo · Bengaluru Area, India

• ✤ Recommended probable, personalized and diverse cart combos which are shown within and cross-restaurants to the customers for faster checkouts and thus minimizing the cart drop off rate.
• ✤ Technologies used: Qubole, PySpark, Presto SQL, AWS, Hive.
• ✤ Performed in-depth data analysis and trained the model with more than 5 billion data points.
• ✤ Devised customized unsupervised evaluation metric for the evaluation of the cart combos.
• ✤ Best Results -> 0.80 AUCROC using gradient boosted trees.

Jul 2017 – Sep 2019 · 2 yrs 2 mos