"Try it Out" Links
All three links lead to projects which are all entirely finished and available to the public. To run either of the github projects, it is necessary to have access to your own personal supabase and claude api accounts. To demonstrate the translation completed in both projects, the site libermormoni.com was created. Enjoy the fruits of our labor!!!! Coloni Adscriptii => AI Agent Collaboration Dashboard and Coordinator Liber Mormonis => Site to view translation produced by Coloni Adscriptii
Inspiration
I was recently speaking with a friend of mine who had served in the Trinidad and Tobago mission. I learned that apparently many people did not have access to the Book of Mormon because it was not translated in the native tongue of those who resided there. While there were some excerpts translated, such as 3 Nephi 11 and 1 Nephi 8, the book in its entirety was inaccessible to them. That is TRAGIC!!!
What it does
We two projects to show for our work. The first is the tool which we used to help ai agents collaborate with each other to peer review scholarly translation. This tool translated over 6600 verses of scripture in a matter of 2 hours. The second project was a website available at the domain libermormoni.com which shows a side by side comparison of the english and latin scriptures.
How we built it
We Implemented the research of one Lawrence Venuti; a translation scholar who pioneered the idea of "literary quality" over all. We collected data to guide the ai agents, (we did translations from a locally hosted LLM as well as Anthropic's Claude) to translate with prosaic beauty and textual fidelity in mind. Each translation went through several rounds of scrutiny and collaboration before it was accepted. We used ReactJS, Vite, Supabase, and Python to build these tools.
Challenges we ran into
- When attempting to run an LLM locally on Braden's machine, we discovered that a connection from his computer to mine would be impossible using his application, LM Studio. As such, we built a python script which I could ping from my computer sending requests that were manually inserted into the LM Studio software. It worked flawlessly! Testing the ai agents without wasting money on expensive tokens is attributable to this!
- 2 hours before we had hoped to finish, we found that the ai agents had stopped returning data in the format which we instructed them to. This was catastrophic. We were quickly running of time; time which we desperately needed to translate the Book of Mormon. Claude Haiku 4.5 struggled to return JSON data without it breaking. To overcome this, we invented our own simpler format to relay data from claude. Instead of the cascading brackets and fickle quotation marks, we structured the data by categories and subcategories. Each category was surrounded by three successive percentage signs: '%%%' and each subcategory was surrounded by three dollar signs: '$$$.' Thus we were able to interpret and refine the data as we pleased.
- The pizza was running out and so we ran downstairs. Disaster averted!
Accomplishments that we're proud of
To translate the entirety of the Book of Mormon into Latin, it was necessary that we accounted for the expenses incurred by heavy token usage. Beyond simply implementing the caching mechanisms provided by Anthropic, we redesigned the peer-review process to save both time and money by a margin of ~33%. This was done by removing the redundancy that was having separate translators and reviewers. The translators became the reviewers and visa versa. What was once potentially 6 api requests before, was reduced to merely 4 simply by removing the wasted agent! Both projects have a README.md with proper documentation great documetation
What we learned
Neither of us had experimented much with locally supported LLMs prior to this point. The struggles of getting it running were particularly frustrating and enlightening. We learned about CORS, as we had to play with it to allow intra-network connections. We played around with the port and streaming mechanisms provided by certain python libraries and learned a great deal about networking. Additionally, I had never worked with ai agents prior to this point. Getting them to work in tandem in the peer review process was a nightmare. This is a rewarding skill gained through painful tinkering.
What's next for The Pen of John Wycliffe
We would like to find 3 different languages which are relatively prominent yet have no translation of the book of mormon. Then it will be necessary to contact people who speak these languages (or either spanish or english) and work with them to create a proper and qualitative translation of the Book of Mormon. It could also be of great benefit to contact the Missionary Department so as to discover where these places may be. The field is white and ready to harvest!!!