30 Years of Records,
One Question Away
How Kodexo Labs built a fully on-premises AI system that lets a US affordable housing nonprofit search decades of mortgages, deeds, leases and scanned files in plain English. Every answer cites its source. Nothing touches the cloud.
Overnight, it reads 30 years of messy files
- Deed of trust
- Lease
- Board minutes
- Intake note
- Loan ledger
- Property
A counselor asks, in plain English
A homeowner wants a cash-out refinance. Can our second mortgage stay in place?
- Where's this recorded mortgage?
- Which forms does the county need?
It answers, with a source for every claim
No. Since 2019, cash-out refinances can't keep the second mortgage in place[1]. That replaced the 2014 rule[2].
- 1Board minutes 2019 · PDFCurrent
- 2Board minutes 2014 · scanReplaced
The records don't say the current loan balance. Check the accounting system.
One click opens the original file
Policy updated: cash-out refinances no longer eligible for subordination.
Project overview
A Housing Nonprofit That Runs on Its Records
Our client is a US nonprofit with a small team that does a great deal of work, and almost all of it depends on records: who owns what, who lent what, and what was promised to which family and when.
They wanted AI their own staff could use every day, without handing those records to an outside AI company.
- 01
Builds affordable homes, so it tracks who owns what.
- 02
Finances affordable homes, so it tracks who lent what.
- 03
Counsels first-time buyers, and tracks what was promised to each family.
- 04
Manages long-term affordability on the properties it touches.
The challenge
Thirty Years of Memory on One Shared Drive
The organization's memory lives on one shared network drive. Many files are scans or handwritten pages that no search box can read.
The Shared Drive
About 30 years- MortgagesScan
- DeedsScan
- LeasesPDF
- Grant agreementsPDF
- Board minutesScan
- Spreadsheets
- Property photos
- Handwritten pagesUnreadable
Many files are scans or handwritten pages that no search box can read.
What the Staff Questionnaire Found
DiscoveryEach person who had managed the drive over the years had used their own filing system.
- 01Folders were duplicated or empty
- 02Names were inconsistent
- 03Forms were updated in one place but not another
Most of an afternoon
To find one recorded mortgage or one old developer fee.
The real risk
It Was Not Lost Time.
It was an answer given from memory that the records did not actually support, given to:
- A homebuyer
- A funder
- An auditor
Why this is harder than it looks
Three Problems, All Handled at Once
A chatbot pointed at a folder would not solve this. Three problems sit on top of each other, and all three had to be handled at once.
- StartThe RecordsDecades of mixed files on one drive
- ResultAnswers Staff Can UseOnly when all three are handled
Technical · What it looks like
Decades of mixed formats: scanned PDFs, handwriting, spreadsheets, photos, duplicates and files with misleading names.
What it demands
A pipeline that reads every format, makes scans searchable and stays current as staff keep saving new files.
Trust · What it looks like
Staff answer to homebuyers, funders and auditors. An AI that guesses, even rarely, gets ignored or trusted blindly.
What it demands
Every answer tied to the exact source file, and a plain "the records don't say" when the evidence is not there.
Security · What it looks like
The records hold personal and financial information about families. Leadership and the outside IT provider ruled out sending any of it to a cloud AI service.
What it demands
The whole system, AI model included, running on hardware inside the organization and passing a formal IT security review.
Three problems sit on top of each other. All three had to be handled at once.
- Technical
- Trust
- Security
Tap a problem to see what it demands
Our approach
A Trust Problem First, a Technology Problem Second
Four principles shaped every decision.
The AI model, the search index and every document run on one server inside the client's organization. No document content goes to any cloud or outside AI provider.
Each answer carries footnotes that open the exact file it came from, so staff verify rather than trust.
The system reads a read-only copy. Staff keep working exactly as they did, and their files cannot be changed or lost by the AI.
The production server was bought at the start, not the end. We built directly on the machine the system runs on, so nothing had to be migrated.
One Team, One Point of Accountability
Our team owned every step below, with every part documented so the client's IT provider can maintain it.
- 1Discovery
- 2Staff questionnaire
- 3Hardware specification
- 4Ingestion and search pipeline
- 5Chat and reporting experience
- 6IT security review answers
- 7Testing
- 8Staff onboarding
The solution
A Private Document Intelligence System
It reads the organization's whole archive and answers staff questions in plain English, with a source behind every claim.
How It Works
8 steps · one in-house server- 01Copies the drive, read-only
- 02Reads every format
- 03Makes scans searchable
- 04Adds new files each night
From any workstation in the organization, through a browser, each with their own account.
"When were the bylaws last updated?"
- 06Finds the evidence, across documents
- 07Uses the newest rule
- 08Cites every claim, or says the records don't
Staff stay in the highlighted step. Everything before it runs on its own overnight, and every answer arrives with the files a person needs to check it.
What Staff See
Illustrative- A
A Plain English Answer
Drawn only from the organization's own records, even when it spans several documents.
- B
A Footnote Beside Every Claim
Click it and the exact file opens, with its folder path.
- C
Newer Evidence Called Out
When a policy or agreement was replaced, the answer cites the current one and shows what it replaced.
- D
"The Records Don't Say" When They Don't
If the evidence is not there, the system says so rather than guessing.
- E
One Click to a Draft
Summaries, report sections and template letters, ready to export to Word or PDF.
System architecture
The Whole System, Inside One Organization
The AI model, the search index and every document run on one server inside the client's organization. Staff reach it from their own workstations, and it reads a copy of the shared drive, never the original.
- Chat and Reporting
- Local AI Model
- Search Index
- File Reader
Adoption and security
Built for Daily Use, and Built to Pass IT Review
A system nobody opens is a failed project, however good the code. So the build was shaped around how a small, busy team actually works.
Staff log in from any workstation in the organization through a browser, each with their own account. Anything saved to the drive today is searchable the next morning, and every staff member now searches the archive from their own login.
- 01
The staff questionnaire gave us the real questions people ask. Those became the acceptance test set before handover.
- Where is a named client's recorded mortgage?
- Which forms do the city and county need?
- When were the bylaws last updated?
- What fee was charged on a past development?
- 02
An administrator adds or removes people, and an activity log shows who asked what.
- 03
The system reports duplicate files, empty folders and inconsistent names to staff. It never deletes or moves anything itself.
- 04
The system cleared the client's full pre-deployment security review before go-live: the AI runs on the organization's own server, data is encrypted, and the client's IT team keeps full control.
Results & impact
From Kickoff to Daily Use in About Six Weeks
To find a record, down from 30 to 90 minutes
Timed staff tasks on the same questions, before and after
Of test questions answered correctly, with the source cited
Acceptance test set built from the staff questionnaire
Of records indexed, including scans and handwritten pages
Files indexed against files on the drive at handover
Of records stay inside the organization. Nothing goes to the internet.
Confirmed in the IT provider's security review
Whole Team, Own Login
Every staff member now searches the archive from their own login.
Every Answer Cited
Every answer carries at least one source link, confirmed by a spot check of the activity log.
Declines Rather Than Guesses
The rest of the test questions were declined rather than guessed.
Records Stay on the Premises
No document content leaves the building, and the original files are never changed.
What this means for your organization
The Knowledge Exists. Only a Few People Can Find It.
It sits in decades of scans and folders that only a few long-serving staff know how to navigate. When those people are busy or gone, answers come from memory.
Most of these share this problem
- Housing nonprofits
- Community land trusts
- Community lenders
- Public housing agencies
The same system carries over to
- 01
Audit and Compliance Readiness
With every figure traced to its source document.
- 02
Homebuyer and Tenant Counseling
Answered from the actual loan and lease terms.
- 03
Grant and Board Reporting
Drafted from the record and cited line by line.
- 04
Policy History
So staff can see what changed, when and why.
Why Kodexo Labs
Where Prototypes Become Production Systems
This system went from kickoff to daily use in about six weeks, and cleared a formal IT security review before go-live.
- 51 Products
- Top-Rated on Clutch
- Inc. 5000 Clients
- PhD-Level Expert Team
- 94% Retention
- 01
Trust First, Technology Second
Every decision started from one question: can staff verify the answer? Citations and "the records don't say" were requirements, not add-ons.
- 02
Private by Design
The whole system, AI model included, runs on hardware inside the client's organization. No document content ever leaves the organization.
- 03
One Accountable Team
Our team owned every step, from discovery and the staff questionnaire to the IT security review and staff onboarding.
- 04
Built to Hand Over
Every part is documented, and the system ships with a plain runbook the client's IT provider can maintain.
Questions we get asked
What Teams Ask Before Going Private
Plain answers about privacy, cost, accuracy and setup, for teams thinking about putting their own records behind a private AI search.
Does any of our data go to the cloud?
No. The AI model, the search index and every document run on one server inside the organization. No document content goes to any cloud or outside AI provider, and the client's IT provider confirmed it in their security review.
Do we have to reorganize our drive first?
No. The system reads a read-only copy of the drive, and staff keep working exactly as they did. It reports duplicate files, empty folders and inconsistent names to staff, but never deletes or moves anything itself.
Can it read scans and handwritten pages?
Yes. The system reads every format on the drive and makes scans and handwritten pages searchable. When one is used in an answer, the footnote opens the exact file, with its folder path.
What happens when the records don't have the answer?
It says so. If the evidence is not there, the system answers "the records don't say" rather than guessing. On the test questions, anything it could not answer with the right citation was declined, not guessed.
How long does it take, and who maintains it?
About six weeks: paid discovery, then design, build, deployment and handover. The index rebuilds from the source files, and the system ships with a plain runbook the client's IT provider can maintain.
Years of Important Files Your Team Can't Find, and Can't Let Leave the Organization?
Let's talk about a private assistant that answers your team's questions from your own files, with every answer showing where it came from, and nothing ever leaving your building.
Related work
