Skip to content
We now build for Pharma & Life Sciences
ByteFlow Technologies
Case 09 · Media & Publishing

ByteFlow News Agent — A Hindi Newsroom Engine That Fails Closed

Open ByteFlow News Agent
8
Tenants on one engine
16 / 24 (was 4 / 24)
Fact-audit pass, blind test
3.2s → 0.8s
Portal LCP
72 → 97
Lighthouse performance
Client

ByteFlow News Agent · Own Product

Industry: Media & Publishing

Year: 2026

Challenge

Small Hindi news portals cannot staff a newsroom around the clock. AI can draft articles, but an unchecked model will publish wrong facts, and an unchecked feed will publish a communally sensitive story without a human ever reading it.

Approach

One multi-tenant engine: news sources go in, Claude writes a Hindi article, a second model audits it against the source, and the article is published into the portal's own admin. Sensitive stories are held for a human, and if the check cannot run, nothing publishes.

Outcome

8 tenants on one engine. In a blind test on 24 real sources, the Claude writer passed the fact audit 16 times against 4 for the previous model. Hosting article images on the portal itself took one portal's LCP from 3.2s to 0.8s.

A small Hindi portal lives or dies on speed: the story has to be up before the next portal's. Most cannot afford a night desk. So they copy, publish late, or publish wrong.

ByteFlow News Agent is our engine for that problem. It is built around one rule: a fast wrong story is worse than a late right one.

How a story moves

StepWhat happens
PickSources come from RSS and news feeds; a fast model picks stories worth covering
WriteClaude Sonnet writes the article in Hindi, with headline, summary and tags
AuditA second model checks every claim in the draft against the source
SensitivityA term list plus an AI check flags communally sensitive stories
PublishThe article goes into the portal's own admin, through one agent login

A portal joins without any code change on its side. The engine logs in like an editor would. It can also edit or remove what it published, and it reads view counts back from each portal.

Failing closed

Sensitive stories are never published automatically. They wait in a queue for a human editor. If the sensitivity check itself cannot run, for any reason, nothing publishes. A broken guard stops the line; it does not wave stories through.

Only articles that pass the fact audit are eligible to publish without an editor. Today, publishing runs with a human pressing the button while we raise originality further.

Measured, not assumed

When we moved the writer to Claude, we ran a blind test on 24 real sources. The audit model judged each article without knowing which model wrote it.

WriterPassed the fact audit
Previous model4 of 24
Claude Sonnet16 of 24

A smaller fix mattered just as much to readers. Each article's image card used to load from our server. Uploading it to the portal itself took one portal from 3.2s to 0.8s LCP and from 72 to 97 in Lighthouse performance. Prompt caching across portals cuts the cost of every repeated instruction.

What carries over to client work

Any AI system that publishes on your behalf needs three things before speed: a second check on facts, a human gate for the stories that can hurt, and a default that stops when a check fails.

Ready When You Are

Start Your Transformation

One 45-minute discovery call. We walk you through scope, timeline, and pricing — no boilerplate.