Digitizing 40 Years of Kenya’s Coastweek Archive to Test and Reduce AI Hallucinations
My name is Linet Amondi and i work at the Technical university of mombasa as an Eresource & Reference Librarian. On a daily basis i see students come into the library looking for resources on the Kenyan Coast for their research work, others turn to AI tools which gives made up stories because our local archives are not part of modern AI datasets.
This project is an initiative at the Technical University of Mombasa(TUM),Kenya. Through this effort, our team aims at turning the university library 40 year collection of Coast week newspapers from 1978 - 2018 onto clean searchable datasets. These publications cover a vast area of knowledge of coastal history currently housed in the Coastweek journals(https://coastweek.tum.ac.ke/),information that AI tools can't easily access or verify, often leading them to invent facts and fake citations.
Our Library holds decades of coastweek issues which document local politics, coastal trade, tourism, port development and community life across mombasa and East Africa. Right now this materials are trapped in static scans and print volumes that students have to search manually and AI models cannot read and cite it reliably
What we will do:
Our Team will scan 5,000 pages of Coastweek editions into searchable text, then organize the articles by date, headline, people, places, and organizations.
20 Student Research Fellows Selected from TUM's 20,000+ students Population will sit with our library team to read the scanned text alongside the original print copies and fix scanning errors.
Using the verified Coastweek stories, we will put together a public set of test questions to check how often AI tools make up answers about Kenyan history and whether they admit when they simply do not know.
Finally We will build a simple search platform where students can type in research questions and get answers linked straight to the exact Coastweek issue, date, and page number.
What we aim to achieve in one year
Subject to our initial technical audit and copyright clearance, our team will focus on building the core digital archive by scanning and indexing up to 5,000 pages from our Coastweek collection with Twenty Student Research Fellows who will sit with library staff to verify these records and test how AI tools respond to them.
At the same time, we will host practical library workshops and data-checking sessions across campus. Our goal is to reach around 2,000 student participations over the year. By month twelve, we will release an open-access AI test dataset and a search tool that links every answer back to the original Coastweek page.
Therefore we are seeking US$100,000 to fund collection preservation which goes beyond preserving old publications, dataset development, student research, AI platform development, technical infrastructure, training and demonstrating how an African university can use its historical knowledge, student talent and research capacity to address one of the biggest challenges facing artificial intelligence today. .
Our student fellows working with librarians will scan 5,000 Coastweek pages and check them by hand against the bound print volumes on our shelves. That checking process is slow on purpose. A digital file only helps a researcher if someone has actually compared it to the original newspaper. During our campus workshops, sign-in sheets will give us a straightforward count of whether we truly reached 2,000 students.
The search prototype has to pass a practical daily test i.e. when a student asks about coastal history, it must pull up a real Coastweek issue and exact page number. Getting the system to say "this record is not in our archive" is actually harder than making it retrieve text. Still, that refusal matters most in a library.Once we publish the test dataset, other researchers can also check our claims directly and Anyone can run an AI model against those records to see if grounding answers in real Coastweek articles actually reduces error rates.
The funds requested will pay for five areas of work that need to run side by side so we can finish within scheduled time.Their is a need to buy scanning hardware and OCR software because our library does not currently have the equipment needed to digitize this 5,000 pages at high speed. a major share of the budget goes toward stipends for our 20 Student Research Fellows because checking text against fragile print newspapers across a full academic year is demanding research work rather than casual volunteering and paying fair stipends keeps them from dropping out mid-project.
Beyond scanning and verification, we need dedicated technical hours funds to clean, tag, and format the records into open datasets for public use. Another portion covers software engineering and server hosting for the AI search prototype since building a tool that admits when evidence is missing requires more custom work than a standard keyword search. Also the budget pays for workshop materials and facilitation to reach 2,000 students across campus.Every part of this budget ties directly to our promised outputs. If we cut student stipends, for example, we lose the hands-on checking that makes the dataset trustworthy in the first place.
The core team is based right inside TUM Library Services where we already manage the university's digital repository, ISO 9001:2015 quality records, and daily library systems. operating a digitization project inside a busy public university library is very different from working in an isolated lab. We have to balance active shelving space, daily student foot traffic, and ongoing reference services but our staff has already handled smaller digitization batches under these exact conditions.
In addition to that we will recruit 20 Student Research Fellows from TUM's student body and Before any scanning starts, the fellows will go through a training program so our quality checks are consistent from week one. concerning the research side, team members Beryl Oyolla, Linet Amondi, and Zablon Mwaramu recently co-authored peer-reviewed work for the KLISC 2026 conference. That experience means our team knows how to deliver more than just raw scanned files, ensuring our datasets, methods, and testing results hold up to real academic scrutiny.
Realistic failures that could slow down this project even though Coastweek founder Shiraz D. Alibhai officially handed over the 40-year collection to TUM in September 2021, we need an internal sign-off on the exact open-access license for releasing bulk AI datasets publicly which review can take longer than expected. Second, some student fellows will inevitably drop out because Checking scanned text against print volumes is slow and repetitive work over a full year which could make them loose 15% to 25% of their time in school.
The third risk is technical because Building a search tool that refuses to answer when evidence is missing is much harder than simply pulling text. If the AI starts mixing plausible guesses with verified Coastweek citations, it will ruin the library's credibility and defeat the whole purpose of the archive but Even if the project falls short of its full goals, the worstcase outcome still leaves behind something useful for TUM. We would end up with a smaller collection of scanned, verified pages stored safely in our internal library repository (coastweek.tum.ac.ke).
Technical University Of Mombasa Library Services is already covering several costs in kind. The university provides librarian staff time, existing digital repository server (coastweek.tum.ac.ke), and direct access to the physical Coastweek volumes on our shelves. In Kenyan public universities that kind of institutional support is ordinary even when cash budgets for digitization equipment and student stipends are unavailable.
There are no bids on this project.
Comments
No comments yet. Sign in to create one!