Scout and Japahub

The problem

Nigerians planning to work or study abroad are targeted with fake job, visa and study offers, and some of those offers are the front end of trafficking. NAPTIP's documented cases show the same patterns repeating: "processing" and "sponsorship" fees, Gulf domestic work aimed at women, "lucrative" jobs in Ghana and Southeast Asia that are not what they claim. The official information that would let someone check an offer exists, but it is scattered across registers in several countries and, in Nigeria, published as PDF lists. So people pay before they check, because checking is hard.

Constraints

• Privacy by design under the Nigeria Data Protection Act 2023: Scout keeps no personal data, and chats are deleted 24 hours after the last message
• Bootstrapped: no funding and no paid roles, a team of nine including volunteers, and production run by the founder
• Official data in awkward forms: the Ministry of Labour lists exist only as dated PDF editions (2018, 2019 and 2020)
• A model that must never give a verdict: a false "safe" could cost someone a fee or a passport, and a false "scam" could cost a real job

Decisions and why

1. Scout withholds verdicts

   No model can know whether an offer is genuine. It can only report what the official sources say and which documented patterns match. So Scout reports four signals (matched to an official source; some things do not line up; could not verify this; matches known scam patterns), each with its citation, labels web search results as web pages rather than verification, and refers people to NAPTIP and the Migrant Resource Centres, who can act. The person decides, and a human rating on Japahub always outranks Scout.

2. Turned the Ministry of Labour's PDF lists into a checkable dataset

   The Federal Ministry of Labour and Employment publishes its lists of licensed private employment agencies as PDFs, one per year. I extracted the 2018, 2019 and 2020 editions into a dataset that records the edition and the date it was retrieved, so Scout can say "found on the 2020 edition" or "not found on the 2020 edition" instead of claiming a currency the source does not have. An agency licensed later will not appear, and Scout says so.

3. Built an evaluation set before trusting a model

   Before trusting the model with real offers I built a test set from documented cases: scam patterns drawn from NAPTIP cases, and genuine offers from licensed agencies, scholarship boards and verified listings. Each case states the expected signals, the sources that must be consulted and the words that must not appear. In the October 2026 run of Scout v4 (30 cases) the screening layer flagged 18 of 18 scam cases with no false alarms on the 12 genuine offers, and the wording gate found no verdict words. The set is what lets me change prompts, sources or models without guessing whether Scout got worse.

4. Domain-driven design and single responsibility

   Japahub is several products in one codebase: the feed and listings, communities, the business platform for advertisers, and Scout. Modelling each as its own bounded context, with modules that have one reason to change, meant one engineer could ship the advertising flow without touching verification, and a small team could change one part without re-testing everything.

Result

Selected for the 2nd ECOWAS Startup Awards (ECOWAS Commission, Abuja, September 2026). Scout v4 screening test, October 2026: 18 of 18 scam cases flagged, 0 false alarms on 12 genuine offers. 150+ registered users and 618 newsletter subscribers as of November 2026.

As of Nov 2026

Stack

• TypeScript
• PostgreSQL
• Prisma Pulse
• Docker
• AWS EC2
• AWS ECR
• S3
• Domain-driven design

Evidence and links

• About Japahub
• Scout demo video