Skip to content
(01 / 05)Work / FactWatchFact-checking · News2025

A newsroom that ships verified truth, faster.

We built the editorial workflow, claim database, and LLM-assisted verification pipeline behind a leading Bangla fact-check organization, turning a multi-day claim cycle into a same-day one, in Bangla and English.

Client
FactWatch
Services
Product · Web · ML
Duration
14 weeks
Team
4 people
Year
2025
FactWatch, homepage
12h
Avg. claim turnaround
94%
Editor approval rate
3.2×
Daily output uplift
14wks
From kickoff to launch
At a glance

The project in one minute.

The short version, for anyone who needs the story before the detail.

Challenge

A multi-day claim cycle was the bottleneck, and disinformation moves in hours.

Solution

One editorial workspace where a claim moves from intake to published verdict, with an assistant that drafts the evidence timeline for an editor to approve.

Result

A multi-day cycle compressed to same-day, without an editor-trust regression.

The brief

A multi-day claim cycle was the bottleneck, and disinformation moves in hours.

Editors were drowning in screenshots, chat-app forwards, and contradictory sources, with tooling that amounted to a shared drive and a spreadsheet. Claims took days to verify; the most viral ones spread long before a fact-check ever published.

They needed one platform that ingested public submissions, helped editors triage and assign, accelerated source-checking with AI, and produced a public claim record in Bangla and English, without compromising the editorial standard their reputation rests on.

What success looked like

  1. 01Publish verified claims the same day they surface
  2. 02Keep every verdict traceable to its sources
  3. 03Serve Bangla and English readers from one newsroom
What we built

Six parts, one product.

Web · CMS · ML pipeline. Everything below shipped in the first release.

  1. 01

    Claim database

    Every claim, source and verdict in one searchable, linkable record.

  2. 02

    Editorial workflow

    Intake, research, review and publish stages with clear ownership.

  3. 03

    Assisted verification

    Drafted evidence timelines that editors accept, edit or reject.

  4. 04

    Bilingual publishing

    Bangla-first articles with English versions from the same record.

  5. 05

    Verdict labels

    Consistent, shareable rating stamps across web and social cards.

  6. 06

    Archive migration

    Years of earlier fact-checks imported with redirects preserved.

The approach

The decisions behind the work.

14 weeksfrom kickoff to launch, with a 4-person senior team.

  1. 01

    Embedded with the editors for a week.

    We watched the actual workflow: inbox triage, source-checking, the back-and-forth with experts. The bottleneck wasn’t writing; it was finding the original claim across five chat apps.

    Discovery
  2. 02

    Claim database first, UI second.

    A schema with claim, source, expert, and verdict as first-class entities. Every surface reads and writes through this single source of truth. No copy-paste between systems.

    Architecture
  3. 03

    LLM as a research intern, not a writer.

    The model pulls candidate sources, flags duplicates against past claims, and drafts the timeline. Editors approve or reject every line. It accelerates the human; it never publishes.

    AI assist
  4. 04

    Bangla-first, English-second.

    Bangla search, paired headlines in both languages, and structured-data markup so each verdict surfaces correctly in Google’s fact-check carousel.

    Bilingual delivery
  5. 05

    Migrated years of archives in a weekend.

    Thousands of historical claims imported, deduplicated, and back-tagged so the platform launched with the credibility of an established archive on day one.

    Launch

Timeline

  1. Weeks 1–2Discovery

    Stakeholder interviews, audit of the existing product and a written scope.

  2. Weeks 3–5Design

    Flows, wireframes and a design system reviewed in weekly sessions.

  3. Weeks 6–12Build

    Two-week sprints with a working demo at the end of each.

  4. Weeks 13–14Launch

    Performance pass, content migration, go-live and hand-over.

Who worked on it

  • Project leadFerdous Rahman
  • DesignZahid Hasan
  • EngineeringMohon Das
  • EngineeringShakil Ahmed
In the wild

See the whole thing.

Each screen is the full page, top to bottom. Open any one to scroll the entire design.

The result

A multi-day cycle compressed to same-day, without an editor-trust regression.

3.2×More verified claims published per editor, per week

Avg. claim cycle
12h
AI-drafted timelines approved
94%
Archive claims migrated, day one
11k
Editorial corrections post-launch
0
MeasureBeforeAfter
Time from claim to verdict2–3 daysSame day
Tools an editor switches between51
Claims published per editor, per weekBaseline3.2×
“They embedded with our editors instead of pitching us. The platform reads like one of us built it, because effectively, one of us did.”
Editor-in-ChiefEditorial lead · FactWatch
Takeaways

What we'd tell another team.

Three things this project taught us that we now bring to every engagement.

  1. 01

    Sit with the people doing the work

    A week inside the newsroom changed the scope more than any workshop would have.

  2. 02

    Assist, never decide

    Editors trusted the assistant because every draft was theirs to reject.

  3. 03

    Migrate early

    Moving the archive in the first sprint surfaced the hard data problems while there was time to fix them.

Services on this project

Built with

  • Next.js
  • TypeScript
  • PostgreSQL
  • Tailwind CSS
  • Bangla NLP
Let's build

Want results like these?

Tell us what you're building. We reply within 24 hours. No commitment, no sales pitch.

Prefer email? hello@codeflee.com · WhatsApp