Wide & Deep Learning
Memorization plus generalization for app-scale recommendations.
Write-up completePersonal projects · Selected work
I make ideas tangible.
I’m Shivani. This is a living collection of projects I’ve shaped, experiments I’ve tried, and things I’m learning along the way.
Projects
05 / 2026Featured · Interactive learning01
Recsys Field Notes
An interactive recommendation-systems learning notebook spanning a 12-paper reading arc, from Wide & Deep to Pinterest’s modern recommendation stack.
4 write-ups complete · 7-day reading tracker
Explore field notesResale · Product02
Trash or Cash
A practical case study for making marketplace listing decisions faster and less fuzzy.
Open projectLearning · Play03
Tamil Together
A family-centered Tamil learning idea shaped around toddler motivation and everyday speech.
Open projectBuild · Experiment04
Little Machines
Small interactive experiments for turning curiosity into something kids can touch and test.
Experiment in progressLittle Explorer · Recsys capstone05
PlayScout
An activity curator concept for discovering the next right project, kit, or outing.
Concept in progressA little about me
Curious by default
I like finding the shape of a problem, making a useful first version, and refining it through real life.
Focus
Thoughtful products, family life, creative technology, and spaces that work beautifully.
Approach
Warm, practical, detail-minded, and always willing to test a surprising idea.
Based
California · Working across time zones
Applications & recruiter-led opportunities
Interview Pipeline
Four recruiter-led opportunities, one still active.
Updated Oct 7, 2026
Total opportunities4Recruiter outreach
Recruiter outreach4100% of pipeline
Technical screens4All recruiter-led
Active1Patreon
Rejected375% of pipeline
Path through the interview process
Band width represents the number of opportunities.
Recruiter-led
Active
Rejected
Company journeys
| Company | Source | Current / furthest stage | Outcome |
|---|---|---|---|
| Expedia | Recruiter outreach | Technical screen | Rejected |
| Gusto | Recruiter outreach | Technical screen | Rejected |
| ZipRecruiter | Recruiter outreach | Full interview loop | Rejected |
| Patreon | Recruiter outreach | Technical phone screen | Active |
Interactive learning · Recommendation systems
Recsys Field Notes
A reading tracker and working notebook for recommendation systems: classic ranking, retrieval, feature crosses, deep retrieval, marketplace and content recommenders, and Pinterest’s modern stack.
12paper reading arc
4write-ups complete
7day reading tracker
7-day reading tracker
Learning sprintDay 1Frame the product questions and baseline recommender vocabulary.
Day 2Wide & Deep, feature crosses, memorization, and generalization.
Day 3Two-tower retrieval and candidate generation tradeoffs.
Day 4Ranking objectives, calibration, and business constraints.
Day 5Sequence models, session intent, and embeddings in motion.
Day 6Pinterest-style visual discovery and graph-informed retrieval.
Day 7Capstone notes: evaluation plan, product risks, and next build.
12-paper arc
4 / 12 completeDeep Neural Networks for YouTube Recommendations
Candidate generation and ranking as separate systems.
Write-up completeNeural Collaborative Filtering
Replacing matrix-factorization assumptions with learned interactions.
Write-up completeDeep Interest Network
Attention over user behavior for ad and commerce ranking.
Write-up completeTwo-Tower Retrieval
Embedding users and items for fast approximate nearest-neighbor recall.
QueuedSequence-Aware Recommendation
Session intent, next-item prediction, and temporal behavior.
QueuedGraph-Based Recommendation
Edges, neighbors, and propagation signals for discovery.
QueuedExploration & Bandits
Balancing short-term relevance with learning and inventory freshness.
QueuedEvaluation & Counterfactuals
Offline metrics, online tests, bias, and feedback loops.
QueuedMarketplace Recsys
Supply, demand, fairness, seller quality, and liquidity constraints.
QueuedVisual Discovery
Multimodal retrieval for taste, style, and intent signals.
QueuedPinterest Modern Stack
From retrieval to ranking to discovery surfaces at product scale.
QueuedCompleted write-ups
Wide & Deep
How explicit crosses and learned embeddings work together.
YouTube DNN
Why retrieval and ranking want different model shapes.
NCF
What neural interaction layers add beyond dot products.
DIN
How attention can make behavior history more context aware.