Ikram Ali

I built a machine learning-powered hotel room ranking and discovery system for KAYAK to improve room matching, search relevance, and room comparison quality. The system helped users find more accurate and affordable room options by categorizing hotel rooms using raw provider data, price signals, and room type information.

I worked on the applied ML and product side, collaborating with product, data, and engineering teams to improve hotel discovery and search quality. The project contributed to a 23% improvement in KAYAK search index ratings, increased user engagement, and generated 7% additional revenue.

I developed review intelligence systems that automatically generated pros, cons, key themes, and topic tags from user reviews across hotels, airlines, and cars. The system helped users quickly understand important review details and improved organic search visibility through concise, keyword-rich content.

I designed and contributed to NLP pipelines that extracted meaningful phrases from review text and converted unstructured review content into searchable, user-friendly summaries. The project improved user engagement and session duration by approximately 35%.

I built an ML-based image tagging system to automatically annotate millions of hotel images using stakeholder-defined categories. The system also included a watermark-detection model to filter out low-quality visuals and improve the quality of hotel image collections.

The image tags improved hotel detail pages, supported SEO, and helped users make better booking decisions using clear visual cues. The project boosted user sessions by 10% and improved conversion rates by 6%.

Built a machine learning-based Translation Confidence Engine to evaluate the quality of translated content across multiple languages. The system acted as a validation layer between translation generation and production use, helping teams verify whether translations generated by internal models, third-party translation services, or LLM-based systems were accurate enough for production.

I designed and developed a Transformer-based regression model using PyTorch to score translation quality between English and target-language outputs. The model used BERT-style encoder representations to compare source and translated text and produce a confidence score. We trained the system using translation data from the data lake and automated the training pipeline with Apache Airflow.

To define production-ready quality thresholds, we evaluated the model using ROC AUC and selected confidence cutoffs that helped separate reliable translations from low-quality outputs. The system achieved approximately 92–93% accuracy and improved the reliability of multilingual content validation by reducing manual review effort and increasing confidence in translation outputs.

Developed an intelligent email parsing and itinerary extraction system for Kayak-style travel booking workflows, where users book hotels through hundreds of external providers. The system extracted structured travel details from provider confirmation emails, including check-in and check-out dates, hotel location, room details, amenities, booking metadata, and other itinerary information. I contributed to the end-to-end ML pipeline, starting with a PySpark and Apache Airflow-based data ingestion and preprocessing workflow for large-scale email data. We built a three-layer machine learning architecture: a transformer-based multi-class provider classification model covering nearly 700 booking providers, a BiLSTM-based sentence classification model to filter ads and irrelevant email content, and a BiLSTM-based Named Entity Recognition model to extract key itinerary entities.

Recently, we enhanced the system with an AI agent pipeline using LangChain, Google SDK, Claude, and OpenAI models to fetch email context, perform feedback-based extraction, parse amenities, convert results into structured JSON, and store them in the required database. Redis was used for memory and contextual state management.

Built a machine learning-powered recommendation engine for an eCommerce platform to personalize product discovery and improve recommendation relevance across multiple product categories. The system supported seven product categories with parent-category relationships and generated recommendations using a combination of users’ historical interactions, real-time category behavior, product metadata, and global engagement signals.

Designed and trained a two-tower neural network model in PyTorch for candidate generation using implicit feedback data such as clicks and browsing behavior. The model learned user and item representations from features including product title, product description, product-specific attributes, category hierarchy, time-of-day signals, interaction history, and popularity-based signals.

Built the training and deployment workflow using AWS SageMaker Pipelines and Docker to support scalable model training, reproducible experimentation, and production deployment. The recommendation pipeline included feature preprocessing, candidate generation, ranking logic, and custom offline evaluation criteria to measure recommendation quality before production release.