Consumer product · Recommendation intelligence

DramasPicks

A catalogue-first drama discovery and community product designed to keep factual identity, human opinion and machine-derived recommendation evidence separate enough to remain trustworthy.

ReactTypeScriptNode.jsPostgreSQLCatalogueCommunityApplied AI

The product idea

Facts tell you what a drama is. People tell you what it feels like.

Most entertainment products collapse very different kinds of information into ratings, popularity or one ranking score. DramasPicks is built around a different separation: catalogue facts have one authority, authored human opinion has another, and machine interpretation must remain traceable back to evidence.

That makes it possible to support natural community conversation and explainable discovery without converting every review, click or dislike into opaque “engagement” truth.

Product model

Catalogue identity, human conversation and intelligence stay distinct.

01Catalogue truth

What a title, person, credit or release fact actually is, with source and identity lineage.

02Human expression

What people wrote, felt, compared, dropped, rewatched, defended or disliked—preserved as authored context.

03Derived evidence

Versioned, provenance-backed interpretation that can remain UNKNOWN when the source does not support a confident claim.

04Personal fit

Recommendation reasoning that combines constraints, caveats and taste context without pretending one universal score exists.

The layers can inform one another, but none is allowed to silently rewrite the others.

Implemented / accepted evidence

A consumer product with unusually serious data boundaries.

Product architecture

Modular monolith

Catalogue, people, accounts, community, evidence, discovery and admin stay as explicit domain boundaries without premature service fragmentation.

Canonical identity

DramasPicks IDs

Provider IDs are mappings, not product identity. Canonical title/person UUIDs survive provider changes and keep community history attached.

Data operations

Staged + guarded

Imports move through normalization, review, promotion receipts, rehearsal and product-visible read checks rather than writing provider rows directly into product truth.

Database model

Beta + rehearsal

Canonical Beta and a separate persistent Owner Rehearsal database are kept distinct, while destructive verification stays disposable.

Community model

Human-first

Reviews, thoughts, discussions, lists, recommendation requests and responses preserve authored meaning, privacy, spoilers and context.

Intelligence

Promotion-gated

Research findings and recommendation evidence live in a separate Intelligence Lab and cannot silently become product ranking truth.

Product proof

A current product surface, not a portfolio mockup.

The screenshot below comes from the current DramasPicks repository visual-evidence suite. I selected the homepage because it demonstrates the product language and information architecture without relying on third-party poster or person imagery.

DramasPicks homepage showing three discovery entry paths and direct title or person search
Current private-beta homepage evidence. Provider-heavy title/person screenshots remain outside this public portfolio until their imagery/rights treatment is deliberately reviewed.

Catalogue engineering

Provider data is input. It is not canonical product identity.

01Sourceprovider records + lineage
02Normalizeprovider-neutral contract
03Reviewidentity / format / lifecycle / rights
04Promotecanonical UUID + receipt
05Verifysearch / browse / title / person

Identity continuity

Provider changes do not rewrite community history.

External IDs live as namespaced mappings. DramasPicks UUIDs remain the stable anchors for titles, people, authored content and evidence.

Current scale work

From 50-title proof cohorts toward generated market acquisition.

Repeated bounded cohorts have proven identity, publication and product read paths. Current catalogue work is moving toward deterministic, resumable selected-market acquisition with guarded promotion rather than one unbounded import.

Database safety

Long-lived product truth and long-lived rehearsal are intentionally different.

DramasPicks uses a canonical Beta PostgreSQL database plus a separate persistent Owner Rehearsal database for realistic browser/dev work. Automated migrations, destructive rebuilds and integration tests use disposable databases instead.

The point is not only protecting production-like data. It is preventing a developer or AI worker from accidentally treating a convenient local copy as canonical truth.

Recommendation intelligence

Research starts with how fans actually express taste.

Research programme

Seven-community longitudinal fandom research

Public fandom communities are used as research microscopes—not as a production corpus to bulk copy or train on without the right boundaries.

authored reviews / thoughtsrecommendation episodespairwise comparisonsnegative constraintsactor / character / performancecontradiction / reversaltaste profile contextlater outcome feedback

Lab owns

research · provenance · derived evidence · benchmarks · engine lineage

Product owns

catalogue truth · user data · product architecture · deployment · release

What I owned

Product judgement across data, community and discovery.

  • Product direction and information architecture
  • Catalogue/domain identity rules
  • Provider/provenance and data-safety boundaries
  • Community/taste/recommendation model decisions
  • Owner browser acceptance and product quality
  • Release/data-operation acceptance criteria

Why this matters professionally

DramasPicks is where I learned to keep product convenience from overriding data truth. The hard parts are not just pages and components—they are identity, provenance, privacy, moderation, provider dependence, human meaning and how machine-derived intelligence is allowed to influence the experience.

Honest boundary

A substantial working product, not a claim of social-network scale.

DramasPicks demonstrates full-stack product engineering, PostgreSQL/domain modelling, catalogue identity, auth/privacy, community architecture, controlled data operations and recommendation-research discipline.

It does not claim millions of users, a shipped Android application, a production fine-tuned recommendation model, or universal metadata rights. Mobile remains a separate pre-implementation direction and intelligence promotion remains gated.

Next product case study

AndhraFresh — transactional supply, pricing, orders and settlement.

Back to selected work