Amar Mandal

Senior Data Scientist with over five years of experience in designing and scaling AI/ML solutions in LLMs, NLP, and computer vision. Awarded Innovation of the Year three times for impactful technology solutions. Proven leader in guiding teams to deliver high-quality, business-impacting results.

• Machine Learning Masters, iNeuron, 04/01/20, 08/01/20
• Deep Learning Masters, iNeuron, 08/01/20, 03/01/21

• English, Full Professional Proficiency
• Hindi, Full Professional Proficiency

Crawler Chef to replace Jr. Developer

• Architected a hybrid extraction platform combining historical regex recipes, API-driven rules, and LLM-generated fallback patterns to automate e-commerce product attribute extraction at scale.
• Built a multi-source data ingestion pipeline across CSV, Google Sheets, MongoDB, and API payloads, with automated crawler-based backfilling for missing labels to improve training and validation coverage.
• Delivered production-ready extraction recipes achieving 99.54% average validation accuracy and 99.72% average match rate on selected candidates.
• Productionized the solution using FastAPI, asynchronous background workers, run-level tracking, and structured event logging to improve auditability, observability, and operational monitoring.
• Established LLM pipeline governance by benchmarking baseline, multi-stage, and comparison workflows, while implementing request-level cost tracking, retry logic, and failure controls.

Parsing via Codex Skills in the Loop

• Designed an AI-assisted rule generation workflow that created extraction logic directly from sample web pages, significantly reducing manual rule authoring effort.
• Built automated validation tests to compare extracted outputs against expected results, improving pre-deployment quality assurance.
• Implemented self-correcting feedback loops where failed test cases triggered automatic rule refinement and re-validation.
• Improved data reliability and consistency for downstream analytics and machine learning pipelines through automated extraction quality controls.
• Reduced debugging effort and increased confidence in production data pipelines by introducing repeatable, test-driven rule generation workflows.

Image Checks for E-commerce website (Client: Kroger)

• Defined and solved a retail catalog consistency problem by building AI-based validation checks to detect image-content mismatches and product quality gaps.
• Scaled the validation pipeline to 500K–1M checks per month across ~250K items, aligned with a biweekly client refresh process.
• Designed decisioning logic to flag non-compliant catalog items and generate AI-driven alternative recommendations to accelerate remediation.
• Productionized the solution as a FastAPI service with asynchronous inference orchestration and MongoDB SHA-based caching, improving latency and reducing inference cost through deduplication and result reuse.
• Enabled higher catalog quality and faster issue resolution for a large retail client through automated validation and recommendation workflows.

Product Matching using Textual Information

• Improved the product matching engine by 20% using textual features such as product name, description, attributes, and price for fashion catalog matching.
• Fine-tuned transformer-based NLP models including ALBERT, DistilBERT, RoBERTa, and XLM on custom retail datasets to optimize semantic product matching performance.
• Took the solution from ideation to deployment, owning experimentation, model selection, API integration, and production rollout.
• Built and deployed the matching service using PyTorch, Sentence-BERT, Hugging Face, FastAPI, Docker, MongoDB, and AWS EC2.
• Strengthened multilingual and semantic matching capabilities for fashion products through custom model training and evaluation.

Visual Crawling

• Built a hybrid visual crawling system to extract information from product PDFs, images, and videos using multi-stage object detection and OCR pipelines.
• Reduced manual QA scraping effort by 70% by automating visual data extraction workflows.
• Used YOLOv5 for object detection and EasyOCR for textual extraction on visual assets.
• Deployed the solution on AWS EC2 to support scalable crawling and automated content extraction.