I build data systems, and I study how systems fail.
I am Ndemafia Wilsmith, a software engineer at Deacons Publishers, where I build the document AI and data pipelines behind the Prep50 study platform. Outside that work, I look for access control flaws in production systems and report them privately.
What I work on
My main work is data engineering for an exam preparation product. That means OCR over scanned textbooks and exam papers, a production question corpus in PostgreSQL, vector search, and LLM classification with checks around it. Most of the difficult work is making failures show up, not getting the normal case to run.
My security research started in October 2024, when I noticed that BillPoint, a payments app I was using, would return another customer’s phone number and wallet balance from their email address. I confirmed it on my own account and reported it privately. In October 2025 I reported a broken access control flaw to Miva Open University, which fixed it and paid a bounty. In December 2025 I reported a flaw in the Corporate Affairs Commission’s document system through ngCERT, which validated it within 48 hours.
I also built and deployed blordgroup.ng for B-Lord Group as sole developer. Finding authorisation bugs in other people’s systems is the main reason I check authorisation early in my own.
- ROLE
- Software Engineer, Deacons Publishers (Prep50), 2025 to present
- FOCUS
- Document AI, data pipelines, access control testing
- DISCLOSED TO
- CAC · Miva Open University · BillPoint
- VALIDATED BY
- ngCERT
- SHIPPED
- blordgroup.ng
- CODE
- github.com/ndemafiawilsmith
- LOCATION
- Remote
MILSET Expo-Sciences International, Abu Dhabi
My secondary school selected me, and covered the cost, to exhibit at MILSET Expo-Sciences International in Abu Dhabi, an international science fair for young scientists.
My project was a carbon dioxide emission reducer for vehicle exhausts. It passed exhaust gas through a chemical treatment that converted the acidic carbon dioxide into a basic compound, so less of it reached the air.
- EVENT
- MILSET ESI 2019
- HOST CITY
- Abu Dhabi, UAE
- PROJECT
- CO₂ emission reducer
- ROLE
- Exhibitor, school funded
From the exhibition floor
ABU DHABI · UAE









The record so far
- 2019
Exhibitor at MILSET Expo-Sciences International, Abu Dhabi
My secondary school selected and funded me to exhibit a chemistry project: a device that reduces carbon dioxide emissions from vehicle exhaust by chemically converting the acidic gas into a basic compound.
- OCT 2024
First disclosure: BillPoint
While using the BillPoint payments app as a customer, I found an account lookup that exposed other users' phone numbers and wallet balances. I confirmed it on my own account only and reported it privately.
- 2025
Joined Deacons Publishers (Prep50) as a Software Engineer
Sole engineer on the data side of an exam preparation platform: OCR, the production question corpus, vector search and LLM classification.
- OCT 2025
Disclosure to Miva Open University
Reported a broken access control flaw exposing student data directly to the university's CTO. It was fixed and a bug bounty was paid.
- DEC 2025
ngCERT validation at the Corporate Affairs Commission
Reported a broken access control flaw in the federal corporate registry's document system to ngCERT, which validated it within 48 hours.
- 2026
102 scanned textbooks turned into a cited, searchable corpus
31,887 pages OCR'd, 83,807 sections linked into a searchable corpus, and 99.3% of sections given a page citation, for $38 in total cash spend.
How I work
Measure before choosing
Thresholds, worker counts and instance sizes come from measurements on the real data. Most of my worst bugs were numbers I carried over without checking.
Make failures visible
A job that reports success while skipping data is worse than one that crashes. I keep failed and empty as different states so nothing is lost quietly.
Report privately first
When I find a flaw, the organisation that owns the risk hears about it before anyone else, with enough detail to fix it.
Hiring, or working on a hard data problem?
I am open to engineering roles and contract work in document AI, data pipelines and web.
Vulnerability reports go to the security address.