Evidence before synthesis
Retain the source URL, extracted material, factual units, capture time, content hash, and source relationship before generating a public article.
Zyel product / Active platform / 2026
Building news intelligence around evidence.
Bold News is a Zyel-created news platform that turns distributed reporting into traceable story records, multi-source context, reader tools, and an operating newsroom system.
Live catalog3,002 published article records
Source network24 enabled feeds across 7 categories
Core platformTypeScript, Node 22, MySQL, static publishing, AWS
The product problem
More news does not automatically create more understanding.
Aggregation can make stories easier to find while preserving the same problems readers already face: repeated headlines, missing source context, unclear evidence, and little visibility into how separate reports relate.
Bold News began as a compact RSS-to-AI publishing pipeline. The larger product opportunity emerged from the evidence already moving through that pipeline: extracted facts, named entities, source records, same-event signals, framing cues, geographic context, and the history of how a story changed.
The challenge was to turn that raw material into a useful reader product without collapsing reporting into a single unexplained AI answer. The system needed to preserve provenance, show uncertainty, support comparison, stay economical to operate, and remain editable by a small newsroom team.
The product thesis
Compare coverage. Interrogate it. Extend it.
The design goal is not to replace reporting. It is to help a reader understand the reporting record and act on what is missing.
Comparison before certainty
Keep separate source records and framing signals visible so agreement, conflict, missing context, and source imbalance are inspectable.
Abstention before invention
Unknown ownership, affiliation, political orientation, or evidentiary support stays unknown until a supported review changes it.
Operation before spectacle
The product includes feeds, moderation, usage cost, deployment, recovery, and mobile delivery—not only the public article interface.
The ingestion system
One path from distributed reporting to a traceable story package.
The production pipeline treats collection, extraction, evidence retention, clustering, generation, media, and publication as separate stages with observable failure points.
Read and select
Load enabled feeds from MySQL or a JSON seed, normalize incoming items, remove duplicates, and select a diverse run within explicit source and article budgets.
Extract the report
Fetch each article and use Mozilla Readability plus paragraph scoring to retain useful text, core facts, metadata, short quotations, and likely named entities.
Preserve provenance
Write a source snapshot containing the original URL, capture time, extracted evidence, typed tags, geography, and a content hash before synthesis occurs.
Cluster the event
Compare report day, normalized title tokens, entities, topics, angles, event fingerprints, actors, actions, quotation hashes, and factual keys.
Merge source evidence
Combine one to three matching reports into a deduplicated fact pack while retaining each outlet, source URL, byline, framing signal, and snapshot relationship.
Draft within constraints
Send the extracted story bits—not an unbounded browsing prompt—through the article-generation service, then validate the structured response.
Create editorial media
Build a factual image brief, generate or reuse an editorial image, create responsive variants, upload assets, and record usage through one provider interface.
Publish and reconnect
Write the article package and rebuild its article, archive, source, tag, entity, market, location, timeline, RSS, JSON, and dependency surfaces.
The evidence model
The article is an output. The evidence graph is the durable asset.
Bold News keeps the records needed to explain where a story came from, why sources were grouped, and which public routes depend on that article.
Source configuration, timestamps, article text, core paragraphs, metadata, quotations, and named-entity candidates.
Retained source captures, claim units, numbers, timelines, context, attribution, uncertainty, and loaded-language observations.
People, organizations, places, topics, actors, actions, event fingerprints, and same-event or follow-up relationships.
The public story plus every source, tag, entity, category, market, and archive surface that must change when the record changes.
This lets the application regenerate static output efficiently, update an existing story when another same-day report arrives, and preserve the path back from a reader-facing claim to its supporting source records.
The reader product
One publishing system, several ways to understand the record.
The public platform is generated for speed but backed by a Node runtime for identity, personalization, comments, requests, contextual data, and mobile APIs.
Article intelligence
Source records, source-balance summaries, framing breakdowns, comments, related clusters, article-weight detail, evidence links, and original-report access.
Structured discovery
Home, archive, category, tag, entity, source, market, location, timeline, coverage-gap, about, contact, and coverage-request routes.
Outlet transparency
A source directory and source pages containing references, observed bylines, article relationships, and ownership or governance fields when supported.
Reader state
Accounts, followed topics, watchlists, portfolio positions, saved alert rules, comments, feedback events, personalized feeds, and story requests.
Local and market signals
Weather and location modules, market groupings, allowlisted quotes, timelines, and nearby-story context connected to the news catalog.
Server-delivered app
A dedicated `/app/` shell and Capacitor wrappers for iOS and Android reuse the same sessions, articles, settings, comments, and personalization APIs.
Bold Lens beta
Turn provenance into an article-level workbench.
Bold Lens is an isolated beta that extends the evidence model into exact annotations, claim-evidence relationships, context gaps, source profiles, and editorial review.
The article remains readable before analysis finishes. A background queue materializes a versioned snapshot, and the interface hydrates only the evidence that is ready. Exact-span cues can identify supported material, primary evidence, timelines, quotations, certainty language, loaded language, and statements needing more support without duplicating the canonical body.
Composable reading overlays
Framing, Sources, and Evidence can be combined while one stable annotation gutter keeps exact flags aligned with the article.
Context and Evidence audit
Supported claims, primary material, needs-more-support findings, missing context, and open questions appear in a predictable hierarchy.
Source Passports
Coverage cards open source context with publisher, timing, reviewed orientation, evidence status, and unknown fields left visibly unknown.
Investigate and Dig Deeper
A reader can turn an entity, claim, or open question into a scoped research request that preserves article and source context.
The operating platform
The newsroom, runtime, and release system are part of the product.
Bold News was engineered to be operable by a small team without hiding the cost or health of its automated work.
Editorial administration
Separate administrator authentication protects feed management, article editing, moderation, story requests, usage reporting, image refresh, product controls, logs, and deployment actions.
Cost visibility
Text and image events flow through a shared AI usage ledger so model, tokens, estimated cost, run context, and administrative reporting use one contract.
Scheduled ingestion
A locked systemd job runs the compiled pipeline on a timer, prevents overlapping runs, records phase metrics, and keeps extraction failures from corrupting shared state.
Release engineering
Versioned bundles, server-side checks, atomic symlink activation, service restarts, public rebuild modes, live smoke tests, and rollback protection form one repeatable AWS release path.
Static speed, dynamic capability
Public editorial pages are prebuilt, while the Node service handles sessions, APIs, comments, personalization, quotes, local context, administration, and mobile bootstrap.
Server-first mobile delivery
Routine interface and feature changes ship through the server; native rebuilds are reserved for Capacitor configuration, plugins, signing, icons, or deep links.
Engineering for scrutiny
Move from a working pipeline to a platform with explicit seams.
As the product expanded, architecture work focused on making ownership, state, external providers, and release behavior inspectable rather than replacing the proven stack.
- `AppRuntime` now composes environment, database, identity, content, operations, and AI services instead of scattering stateful globals.
- Named data wrappers own users, sessions, comments, devices, published articles, source snapshots, usage events, releases, and pipeline state.
- Text and image generation depend on provider interfaces, allowing production providers, placeholders, usage recording, and tests to share a stable contract.
- Route families are being extracted behind injected runtime contexts and shared HTTP primitives rather than growing one unbounded server file.
- Type checking, linting, formatting, documentation checks, unit tests, output verification, dependency audit, build benchmarks, and browser tests form the validation surface.
- Generated pages carry a dependency manifest so an article change can identify the article, tag, entity, category, source, market, and shared routes that need rebuilding.
What changed
A small aggregator became an evidence-aware operating system for news.
The meaningful outcome is not simply more generated pages. It is a product shape capable of retaining evidence, comparing coverage, supporting reader state, governing AI work, and shipping safely.
- Multi-source stories preserve their individual reporting records rather than erasing provenance during synthesis.
- Readers can move from an article into its sources, entities, tags, coverage relationships, local context, and structured requests.
- Editors can manage feeds, content, images, comments, requests, usage, logs, and releases through protected operational surfaces.
- Static publishing keeps the public site fast while API-backed identity, personalization, and mobile capabilities remain available.
- The isolated Bold Lens path allows evidence and scoring work to mature without presenting provisional analysis as settled fact.
Why it matters
AI products become credible in the systems around the model.
The differentiating work is not the API call. It is the evidence contract, source retention, failure behavior, human review, cost visibility, interface, data model, deployment path, and explicit boundary between known and provisional.
Bold News demonstrates Zyel's ability to shape a product thesis, build the data and application architecture behind it, create the public and operational experiences, and carry the system through infrastructure, release, measurement, and continued iteration.