I turn collections nobody can search (newspapers, manuscripts, records, film) into data you can question, at the scale of hundreds of thousands of pages. Every answer points back to the page it came from.
Email Jamesor call 305 781 2280
capture logR031 · F0499
106,393 pages captured · 290.6M words
How were public lands sold, and how did yerba mate grow after the war?
The papers show the state did not sell the yerbales on the open market. It kept them as public property and leased them. Yerba exports grew fivefold, concentrated in a few companies.
Real answer · 12 passages · 5 papers · 1877–1897
One page, twice: as the scanner left it, and as the pipeline reads it, with every article, ad and table found.
The fields every page gets. A page that comes back short, empty or invalid moves up a tier.
Eight steps, from a fragile page to a cited answer. Pages enter once and are read by the cheapest model that can handle them; hard pages escalate. Every search runs keyword and meaning together, then answers only from cited pages.
Track down the surviving copies. Give every page a stable ID, a master scan and a checksum.
stable ID · TIFF + JPEG master · checksum
Six models raced on five of the hardest pages. No model won every page.
370 saved results across 34 runs
A one-line request became an editorial spec: keep the printed wording, flag doubtful words.
inspect · segment · transcribe · validate
A fast model reads every page first; only the pages it fails move up a tier.
under 50 words, invalid JSON or empty: next tier
Pages become overlapping passages stored by meaning, so different words for one idea meet.
ferrocarril · camino de hierro · railway · transporte
Keyword search finds exact names; meaning search finds ideas. A second model reranks both.
merge · filter · balance · rerank
The best passages form an evidence pack. The answer may use only what it cites.
Quick 12 · Deep 30+ · Full: all passages
Every claim links to its paper, date and page. The original scan is one click away.
La Democracia
1881 · R031 · F0499
The overview is above. The full technical build, step by step, arrives as a one-time link sent to the address you enter.
Contents8 parts
Sealed · opens with a one-time link
One email, no password. I see who reads the full build; nothing else is shared.
Every step above, in full: what was tested, what was chosen and how the parts fit together.
Ingestonce per collection
Queryevery search
source + preserve
The first problem was historical, not technical: finding where the papers still existed, then turning fragile pages into protected digital objects.
scattered sources
one protected record
model selection
No two pages share a layout. Plain OCR reads lines and mixes columns. A multimodal model reads the page structure first. Six models went head to head on five of the hardest pages.
| Model | Score, 5 hard pages | Valid JSON | Avg time |
|---|---|---|---|
| Advanced model Achosen | 93.6% | Valid JSON | avg122 s |
| Advanced model B | 65.7% | Valid JSON | avg174 s |
| Fast model C | 58.9% | Valid JSON | avg83 s |
| Fast model D | 57.8% | Valid JSON | avg169 s |
| Small model E | 56.4% | Valid JSON | avg231 s |
| Small model F | 25.7% | Valid JSON | avg138 s |
Model names withheld. Squares: valid JSON out of 5 runs. Time: average per page.370 saved results · 34 documented runs
Hard pages, with what makes each one hard.
prompt testing
Small wording changes moved coverage. Dozens of experiments turned a one-line request into an editorial specification.
“Extract the text”, then “Transcribe exactly”, then inspect · segment · transcribe · validate
the four rules
batch pipeline
Running the best model on every page is slow and expensive. A tiered factory sends each page through the cheapest model that can handle it.
Fast first read of every page, temperature zero, structured JSON.
escalate if < 50 words, invalid JSON or empty output
Dense, damaged or sparse pages. Known-hard papers such as La Democracia go here directly.
escalate if < 50 words, invalid JSON or empty output
Output limits, recitation blocks, empty answers. Handled by observed pattern, page by page.
embeddings + index
Words change over 160 years. Each page is split into overlapping passages and each passage becomes a vector stored in PostgreSQL with pgvector.
overlapping passages
one idea, many words, 160 years
hybrid search + rerank
Keyword search alone misses meaning. Vector search alone misses exact names. Running both, then reranking, finds the right passage across tens of thousands of pages.
keyword
Exact names, places, prices. Finds “Biltmore” and “R031”.
vector
Meaning. Finds “land sales” when the paper says “arriendo de yerbales”.
evidence packet · best passages with source links
search prompt + answer
This is a research tool, not a chatbot. The best passages form an evidence pack, and the answering model writes under strict rules.
Question, then Evidence pack, then Answer model, constrained, then Answer
search prompt rules
depth
search UI + results
The technology disappears behind a place to ask, browse and check. The original scan is always one click away.
How were public lands sold, and how did yerba mate grow after the war?
The papers show the state did not sell the yerbales on the open market. It kept them as public property and leased them. Yerba exports grew fivefold, concentrated in a few companies.
Real answer · 12 passages · 5 papers · 1877–1897
cited scanR031 · F0499
image search
Search what a photo shows, not just its caption. On coralgableshistory.net, try “house on fire” or “Venetian Pool”.
¿Cómo se vendieron las tierras públicas y creció la yerba mate tras la guerra?
Los periódicos muestran que el Estado no vendió los yerbales en mercado abierto: los mantuvo como propiedad fiscal y los arrendó. La exportación de yerba mate se quintuplicó, concentrada en pocas empresas.
La Democracia · 1881
La Reforma · 1877
La República · 1891
El TiempoReal answer · every claim links to its paper, date and page
Every model read the same five pages. One advanced model was the most stable; no model won every page. Model names are withheld.
| Model | Score, 5 hard pages | Valid JSON | Avg time |
|---|---|---|---|
| Advanced model Achosen | 93.6% | Valid JSON | avg122 s |
| Advanced model B | 65.7% | Valid JSON | avg174 s |
| Fast model C | 58.9% | Valid JSON | avg83 s |
| Fast model D | 57.8% | Valid JSON | avg169 s |
| Small model E | 56.4% | Valid JSON | avg231 s |
| Small model F | 25.7% | Valid JSON | avg138 s |
Squares: valid JSON out of 5 runs. Time: average per page.370 saved results · 34 documented runs
The five pages every model had to read.
Both were built end to end and both are public. Search them yourself.
Paraguay’s press, 1845–1904: every surviving newspaper, many the only known copy.
Five Coral Gables weeklies, 1941–1977. Search a family name, a street or what a photo shows.
Venetian Pool
“…recreation facilities including the municipal golf course, tennis courts and the Venetian Pool.”
“I thank former Ambassador James Cason and his family for the invaluable project. This technology-driven historical archive will reach our schools as an innovative tool for students to explore our past interactively.”
“Through AI this has life and can be used. Had it stayed in an archive in a traditional format, no one would have seen it.”
capture session4 steps
A few pages, a reel, or a list of what’s in the boxes.
What search could find, what’s hard, what it would take.
A slice of your collection made searchable, so you can test it with your own questions.
Every page captured, read and indexed, each with a stable ID and citations.
I built the AI pipeline behind paraguayhistory.net and coralgableshistory.net.
linkedin.com/in/jamescasonMiami, Florida