AI/ML Candidate Evaluation Template
CANDIDATE INFORMATION
• Name: Satpal
• GitHub: [Add if available]
• Portfolio/Website: [Add if available]
• Years of AI/ML Experience: 5 years (applied AI/LLM projects at Amazon)
SECTION 1: AI/ML PROJECTS PORTFOLIO
Project 1: Rufus – RAG Pipeline for Size-Fit Compatibility
• Duration: 2024-2025
• Role: PM II (L5) (Tech-driven CX innovation)
• Team Size: 8 (Engineers, Applied Scientists, Research Analyst, Business Analyst ,
Subject Matter Experts , Sr. Associates )
• Industry/Domain: E-commerce, Retail AI
Problem Statement:
Customers faced high return rates and confusion due to missing or inconsistent size-fit
compatibility details across Amazon’s product catalog. Needed an AI solution to dynamically
answer “Does this product fit me/my needs?” using structured + unstructured catalog data.
Technical Approach:
• Designed a Retrieval-Augmented Generation (RAG) pipeline customized for size-
fit queries.
• Developed domain-specific ANN indexes for compatibility (vs. universal indexes).
• Embedded catalog attributes, customer Q&A, reviews, and return comments into
vector DB.
• Orchestrated LLM (Trex-66B) over structured + retrieved embeddings.
Architecture/Pipeline:
1. Data ingestion: ANDES tables, PDP attributes, UGC.
2. Embedding store: ANN indexes tuned for size/fit queries.
3. Retrieval pipeline: Hybrid keyword + ANN.
4. LLM layer: Trex-66B formats outputs with natural explanations.
5. Feedback loop: PIR reduction & user clicks → retraining.
Key Challenges & Solutions:
• Challenge: Fragmented attributes across multiple catalog tables.
Solution: Normalization + LLM-based attribute clustering.
• Challenge: High latency in retrieval.
Solution: Size-fit optimized ANN indexes.
• Challenge: Hallucinations in LLM responses.
Solution: Grounding via RAG + dynamic guardrails.
Results & Impact:
• 20.7% of returns tied to size-fit → Rufus targeted 13–15% reduction.
• Reduced PIR (Purchase Inquiry Rate) on test PTs by ~11%.
• Projected $1.4B NCRC impact WW if scaled.
Code Repository: [Internal Amazon Git]
Demo/Deployment: [Amazon internal sandbox – Rufus]
Documentation: [Working Backwards PRFAQ, Design Docs]
Project 2: AI-Assisted Buying Guides
• Duration: 2024 – ongoing
• Role: PM II ( L5) (Lead, Size & Fit Customer Experience)
• Team Size: 8 (Engineers, UX Designers, Research Analysts, Applied Scientist ,
SMEs, Business Analysts)
• Industry/Domain: E-commerce, GenAI CX
Problem Statement:
Shoppers evaluating complex products (e.g., jeans, coats, headphones, luggage) struggled
with dense attribute data. Existing buying guides were static and generic. Needed
personalized, AI-driven buying guides to improve decision-making and reduce returns.
Technical Approach:
• Built LLM-powered buying guides using catalog metadata, reviews, and survey
insights.
• Personalized guides based on customer intent (travel, gifting, fit preference).
• Integrated GenAI tooltips & comparison widgets directly in PDP.
Architecture/Pipeline:
1. Data sources: Structured catalog + RAI dashboard + survey insights.
2. LLM layer: Generates tailored, natural-language guides.
3. UX integration: Tool-tip infographics, comparison tables, AR modules.
4. Feedback loop: CTR, PIR, and conversion uplift data retrains models.
Key Challenges & Solutions:
• Challenge: Selecting relevant attributes from 750+ PT-Attribute pairs.
Solution: Attribute importance scoring (RAI + LLM clustering).
• Challenge: Localizing guides across 6+ marketplaces.
Solution: LLM-based multilingual translation with human QA.
• Challenge: Avoiding generic suggestions.
Solution: Context-aware personalization using customer signals.
Results & Impact:
• Tool-tip launch cut PIR from 27% → 13.5% on pilot PTs.
• Estimated $105M i-GMS uplift across 6 marketplaces.
• Increased evaluation-to-purchase conversion by ~9%.
Code Repository: [Internal Amazon Git]
Demo/Deployment: Deployed on Amazon PDP (Tool-Tip + Guide widget)
Documentation: [AI Buying Guide PRFAQ, Experiment Results]
SECTION 2: RESEARCH CONTRIBUTIONS
(Not applicable – no external research papers published. Could highlight Amazon internal
whitepapers if needed.)
SECTION 3: CODE ARTIFACTS & TECHNICAL
CONTRIBUTIONS
• GitHub Repository: [Private Amazon Repos – not public]
• Languages & Stack: Python, FAISS/ANN, AWS S3, DynamoDB, LLM APIs
• Key Features: Custom RAG pipeline, multilingual buying guide generator, attribute
clustering, ANN optimization.
• Performance Optimizations: Latency reduction with size-fit ANN indexes,
guardrails for hallucination prevention.
• Community Engagement: Internal Amazon collaborations with RBS Tech, Legal,
UXD, CAM teams, Category Team.