CASE STUDY / 2026PRODUCT + BACKEND + UI

TWIZZ.

A private editorial intelligence system that turns noisy technical discovery into thoughtful, copy-ready writing—without handing publishing control to automation.

ROLEProduct & Backend Engineer

STACKPython / FastAPI / SQLite / Jinja

MODELocal-first editorial workflow

Twizz editorial desk showing pipeline activity and a proposal ready for review — 01 Editorial desk — orchestration, review, confidence and approval in one surface.

01 // The problem

Discovery is easy.
Judgment is the work.

Technical feeds produce an endless stream of repositories, papers, discussions, and releases. Saving links is not the hard part. The hard part is deciding what is relevant, what is actually novel, what can be verified, and what deserves to become a public idea.

Twizz was designed as an editorial system rather than an auto-posting bot. It handles collection, scoring, drafting, and organization, then pauses for human review. Every source remains visible and every published sentence stays intentional.

02 // System design

An eight-stage editorial pipeline.

Each stage has one job, a visible output, and a clear hand-off to the next decision.

Discover

Collect candidates from RSS, Hacker News, arXiv, GitHub and manual sources.

Normalize

Turn inconsistent source payloads into one dependable candidate model.

Deduplicate

Prevent repeated URLs and previously processed ideas from entering the queue.

Score

Measure relevance and novelty while retaining low-scoring discoveries for inspection.

Verify

Attach confidence, evidence notes, uncertainty and content flags to each draft.

Draft

Generate voice-aware editorial copy grounded in the bounded source context.

Chunk

Split complete drafts into numbered, copy-ready posts under X limits.

Review

Approve, edit, reject, save or delete—nothing publishes automatically.

03 // Product surfaces

One workflow, three focused views.

Twizz discovery library with filters and source cards — LIBRARY / 02
Everything stays findable.
Search, filter, sort, inspect scoring state and distinguish raw discoveries from proposals without losing source context.

Twizz published archive showing an approved technical thread — ARCHIVE / 03
A record of the final voice.
Approved and edited work becomes a clean body of writing that can inform future editorial decisions.

04 // Engineering decisions

Built for control, not spectacle.

Local-first storage

SQLite keeps the editorial database portable, inspectable, and independent of a hosted service.

Soft flags over silent deletion

Uncertain, sensitive, or low-confidence content stays visible with context instead of disappearing from the system.

Human approval as architecture

The boundary is deliberate: Twizz can recommend and prepare, but the creator owns publication.

Copy-ready by design

Drafts become individual character-counted posts, making the final handoff practical rather than merely presentational.

05 // Outcome

From a feed of links to a deliberate body of work.

Twizz combines backend orchestration, editorial safeguards, transparent scoring, source preservation, and a responsive review interface in one complete local product.

Return to selected work