Muzna Naeem. Machine Learning Engineer

An AI assistant that lets non-technical teams manage product catalogs through natural conversation, eliminating engineering tickets for routine catalog work and giving domain teams direct control of their data.

• Two-tier agent design: a lightweight model handles intent and routing while a stronger model executes the actual operations, with backend tools connected through standardized interfaces
• Smart file ingestion handles the messy real-world formats teams actually send (Excel, CSV, JSON, PDF), with built-in pagination to keep large catalogs from overwhelming model context limits
• Every change passes through a human approval gate, with an automated evaluation harness continuously scoring agent decision quality across a 31-scenario test suite

A multimodal AI service that turns product images into both descriptive copy and structured catalog data, automating the manual data-entry work that used to take hours per product.

• Schema-flexible by design: each product team defines their own attribute structure at runtime, with automatic validation guaranteeing parseable output every time
• Multimodal: takes a product image plus an optional text prompt and returns both narrative copy and structured attribute values in a single response
• Real-time streaming with persistent conversation history, so teams can refine outputs through follow-up prompts rather than starting from scratch

A research effort to test whether fine-tuned embedding models could meaningfully improve search relevance, while building a rigorous evaluation framework along the way that the team could keep using.

• Six-stage evaluation framework measuring ranking quality, result diversity, catalog coverage, and how often users click results that rank high, going beyond a single accuracy number
• Tested two fine-tuning approaches in parallel (one trained on domain-specific synonyms, the other on real user click data) across four different ways of encoding product text
• Statistical significance testing stratified by query type, validated against the industry-standard ESCI benchmark, making results defensible rather than anecdotal

A recommendation engine that learns from how users browse in real time, predicting what each user is most likely to look at next based on their session so far, not just their long-term history.

• End-to-end pipeline using sequential transformer modeling to capture how user interest evolves across a browsing session, with GPU-accelerated feature engineering for fast retraining
• Evaluated on the metrics that actually matter for recommendations (Recall, NDCG, and Mean Reciprocal Rank) before any promotion to production
• Lazy-loaded model artifacts and clean dependency injection keep inference latency low even when serving several models from one service

A search ranking system that learns relevance from real user clicks, replacing static keyword scoring with rankings personalized to actual browsing behavior, so users see what they're likely to engage with rather than what a keyword algorithm thinks they want.

• Modular pipeline with separate stages for feature engineering, model training, and serving, each independently deployable for clean ownership boundaries
• Gradient-boosted ranking models combine click-stream signals with classical text-relevance features to learn what users actually engage with, not just what matches their query string
• In-memory model cache with timed refresh keeps inference latency consistent, while async logging captures every ranking decision in the background for offline analysis

A recommendation engine that learns which products customers commonly purchase together, powering "frequently bought with" suggestions on product pages, cart, and post-purchase flows.

• Built on collaborative filtering over anonymized purchase history, learning the strength of association between every product pair without exposing individual users
• Variant-aware recommendations at the SKU level, important for products that come in multiple sizes, colors, or styles, where averaging across variants loses the signal
• Pre-computed lookup tables keep API responses in single-digit milliseconds, with incremental retraining as new purchase signals come in

A semantic-similarity engine that surfaces comparable products for "you may also like" placements, out-of-stock alternatives, and substitute discovery, matching items on the meaning of their text content rather than just keywords.

• Text embeddings of product titles and descriptions indexed in a vector database for sub-millisecond similarity search across the full catalog
• Configurable strategies for what gets embedded (title only, description only, combined, weighted), chosen per use case based on offline evaluation against historical engagement
• Hot-swap of the index on retrain keeps the service fully available throughout catalog refreshes, with no cold-start latency for end users

A scoring service that quantifies how popular each product is based on real customer engagement, feeding downstream ranking, merchandising, and recommendation systems with a single trusted signal.

• Daily aggregation combines engagement signals (views, clicks, conversions, dwell time) with configurable weights that can be tuned per use case
• Configuration-driven design lets teams adjust the scoring strategy without redeploying, supporting rapid iteration on what "popular" should mean in different contexts
• GitOps deployment across environments, with zero-downtime rollouts driven entirely from version control rather than manual operator commands

A monitoring service that catches data quality issues before they propagate downstream, ensuring ML pipelines always run on data they can trust rather than failing silently on bad input.

• Multi-layer validation across schema conformance, completeness, document counts, and domain-specific business rules, covering the data layer that ML pipelines depend on
• Dual reporting: real-time Slack alerts when something breaks, plus structured BigQuery metrics for trend analysis and dashboards
• Configuration-driven validation rules let teams add new quality checks without modifying core service code, keeping the engine reusable as new pipelines come online