• Developed & deployed a MMX solution which manages $230M/yr. across 17 channels
• Uses a combination of ridge regression-based attribution model and non-linear optimization
• 2.4% lift in activations compared to BAU allocation–reduced SAC by 4.8% .

• Built the DCO solution for serving dynamic creatives to prospects, optimizing elements in realtime
• Developed the ML web service using AWS Lambda, DynamoDB, API Gateway & CloudFront – which stitches and delivers creatives (HTML/JS Zip bundle) to ad serving platform
• Implemented Bayesian Bandits for continuous optimization of elements towards higher CTR
• ~39% CTR uplift, 0.9% acquisition uplift, and savings of 290 hrs./qtr. Effort

• Developed a chatbot using RAG and LangChain for querying Confluence documentation of tech teams; uses LLama2 7B as the base model, ChromaDB vector store and FAISS for similarity search
• Experimented with PEFT (LoRA/QLoRA) based fine-tuning - but no significant improvement over RAG
• Currently finetuning on Mixtral MoE 8x7B Model using Direct Preference Optimization (DPO).

• Developed the Smart Bidding Algorithm which optimizes for spend on high ROAS patterns (‐17%SAC)
• Led MTA (Multi Touch Attribution) initiative to enable holistic view of channel ROAS
• Built subscription & viewership forecasting model > led to +0.12 AUC for downstream churn models

Lead a team of 6 analysts in a $1.3MM engagement to build a sales force planning and optimization tool to maximize sales uplift or ROI and reduce expenditure. We have clustered outlets, measured impact of visits using a regression model and designed a Mixed Integer programming (MIP) based optimizer to generate visit plans. We have also designed a planning & reporting tool using PowerApps to design plans and report results. Potential impact of ~3% incremental sales uplift(+$6.2MM) and sales force expenditure reduced by 24%($0.5MM) – and savings of 960 hours/quarter effort.

Working for a leading CPG manufacturer, we created a sales driver model for attribution of sales volume to key imperatives using ElasticNet regression. We developed a Trade Promotion Optimization (TPO) framework which enabled the Account teams to design optimized promo plans leading to reduction of ~2100 hrs. of manual effort/quarter. The framework is powered by a linear optimizer on the sales driver model & builds a 52-week calendar detailing optimal price points, ideal promotion weeks, and optimal execution distribution points

Developed a Competitor Clash Forecasting framework to help account managers design promotions in accordance with forecasted competitor prices. The framework uses ARIMAX models to forecast the prices of competitor products and uses these prices to predict manufacturer and category sales through a regression-based sales attribution model. Captured a 30% reduction in negotiation time and sales uplift of 2%($3.6MM) across four retailers in six months

Implemented CNN based YOLO object detection algorithm for a video occasion detection tool – to detect the class and number of objects present in a scene with a mAP of 51.5%. Additionally, we also had to design a XGBoost model to classify a scene into consumption, transaction or celebration based on the class and number of objects detected by YOLO. This was an industry-first dataset for Consumer Behavior Data and acquired by a Global Beverage Giant for ~$5MM

Designed a Neo4j graph database based digital fingerprinting model which identifies unique users across anonymous devices leveraging behavioral relationships. The ideas was to classify new anonymous user IDs into either a current user or "new-to-platform" users. We reduced advertising expenses by $0.8MM/month by reducing number of targeted users