Skip to content

Showcase · AI lab

Four AI patterns, running live

These are interactive simulations of AI systems we build in production. The retrieval, ranking and matching math below runs entirely in your browser — no API keys, nothing you type leaves the page. In client work, the same interfaces sit in front of real embeddings and LLMs.

01Hybrid RAG retrieval

Ask a question against a product knowledge base. Watch keyword search (BM25-style) and semantic search score every passage, blend them with the fusion slider, and get a cited answer from the top hits.

Generated answer

Ask a question to run retrieval and generate a cited answer.

Pure keyword misses paraphrases (“is my data safe” never says “encryption”). Pure semantic can miss exact terms. Hybrid takes both.

Retrieval scores — 8 passages
  • Plans and billingKB-104
  • Data encryptionKB-211
  • Slack and email integrationsKB-118
  • Importing your dataKB-093
  • REST API and webhooksKB-156
  • SSO and access controlKB-201
  • Mobile appsKB-142
  • Support and SLAsKB-088

02Movie recommender

Content-based filtering: like a few movies and the recommender builds your taste vector from genres and vibes, then ranks everything else by cosine similarity — with an explanation for every pick.

Catalog — like at least two
Your recommendations

Like a couple of movies and the recommender starts ranking the rest of the catalog.

Every movie is a vector over 26 genre/vibe features. Your taste profile is the average of what you liked; matches are ranked by cosine similarity, recomputed on every click.

03Resume screener

One job description, any resume. The screener extracts skills and experience, checks them against the role's requirements, and returns a weighted score with exactly what matched and what's missing.

Job description

Senior Frontend Engineer

4+ years · owns UI architecture on a product with real users

ReactTypeScriptNext.jsCSS / TailwindTestingAccessibilityGraphQL (nice)Node.js (nice)Design systems (nice)CI/CD (nice)Performance (nice)
Resume
Screening report

Pick a sample resume (or paste your own) and run the screen.

04Personalized news curator

Pick your city and interests. The curator scores a pool of stories on topic match, local relevance and freshness, and assembles a front page that's yours — with the 'why' shown on every card.

1 · Your city
2 · Interests (pick at least two)
Each story is scored as 55% topic match + 30% local relevance + 15% freshness. Stories bound to other cities are filtered out entirely — a Delhi transit story shouldn't reach a Tokyo reader, no matter how good the topic match is.
Your front page

Pick a city to anchor the local coverage.

In production these patterns run on real models — pgvector or Pinecone for retrieval, Claude for generation, feature stores for recommendations. The demos above keep the exact same UX contract, which is the part users actually feel.