Project manager’s account
“We placed the equipment order after it was approved on June 18.”
PRIVATE AI & INFORMATION DISCOVERY
Find what matters.
Keep it in
your building.
CloseVector builds a private AI system that turns a mountain of records into a clear, source-linked picture: what supports your position, what threatens it, what’s missing and what to ask for next. It runs on hardware you own, in your building, so confidential files and questions stay there.
Private search · Evidence investigation · Custom implementation
A private AI system your team can put to work.
Point it at your records and your question. It helps your team build one source-linked review: the strongest records, the bad facts, the timeline, what’s missing and what to ask for next.
What you get
- The system, configured for your records and the questions you need answered.
- Installed on hardware you own.
- Training, documentation and an agreed support arrangement.
- A written proposal, before work begins, that sets out what your system will produce.
You get a working system for your records, not another AI toolkit to assemble.
What the implementation includesFrom scattered records to a review your team can use.
A company is reviewing a $186,400 equipment order after the units fail testing. The team needs to know what authorized the purchase and which units the supplier agreed to replace.
The illustration below shows the source-linked review CloseVector helps the team build from those records.
Select a source beneath any finding to see the record behind it.
Before the next payment: The team needs to resolve two questions: what authorized the order, and which failed units the supplier agreed to replace.
What the records show
The production units and evaluation units are listed under warranty W-2. The replacement terms are missing.
Test log, July 9: the production units failed the same test as the evaluation units.
What needs explaining
Supplier email, July 15: “We will cover replacement costs for the two evaluation units only. The production order is excluded.”
PO 7316 was placed June 12. The project manager says it followed approval on June 18.
Timeline and people
- :
Purchasing places PO 7316, $186,400.
- :
The approval date the project manager cites.
- :
The production units fail testing.
- :
The supplier’s account manager limits replacement to the evaluation units.
What’s missing
No approval record was found in the reviewed files. The approvals workbook could not be opened.
approvals-q2.xlsx: password-protected, could not open.
Next requests
Any approval record for PO 7316, and the full warranty terms.
An open copy of approvals-q2.xlsx.
Ask the project manager: what authorized the June 12 order?
View the example records
Warranty scheduleAttachment to PO 7316
| Units | Warranty reference |
|---|---|
| Evaluation units EV-1 and EV-2 | W-2 |
| Production units on PO 7316 | W-2 |
Attachment list: warranty schedule received; W-2 terms document not supplied.
Return to the review ↑Test logJuly 9 · Acceptance test AT-3
| Units tested | Test | Result |
|---|---|---|
| EV-1 and EV-2 | AT-3 | FAIL |
| Production batch for PO 7316 | AT-3 | FAIL |
Supplier emailJuly 15 · Account manager → Purchasing
“We will cover replacement costs for the two evaluation units only. The production order is excluded.”
Return to the review ↑Purchase order 7316Issued by Purchasing
| Field | Entry |
|---|---|
| Purchase order | 7316 |
| Order date | June 12 |
| Total | $186,400.00 |
Project manager’s accountStatement supplied for the review
“We placed the equipment order after it was approved on June 18.”
Return to the review ↑Collection coverageSearch record · readable files only
| Search | Result |
|---|---|
| Any approval of PO 7316 dated on or before June 12 | 0 matches |
| Readable files searched | Purchase order, warranty schedule, test log, supplier email and project manager’s statement |
| Excluded from search | approvals-q2.xlsx, could not open |
Intake exceptionsProcessing record
| File | Status | Reason |
|---|---|---|
| approvals-q2.xlsx | Not opened | Password-protected |
What your team can do next
The team can see the conflicting dates, check the supplier’s exact words and identify the records still needed. They can ask specific questions with the supporting documents in front of them.
Why it matters
- Find the record that changes the decision before you commit.
- Learn early whether the records support your position. Sometimes the most valuable answer is that they don’t.
- Explain where the evidence stands, to a client or a colleague, with sources they can check.
Published evidence
How capable can local AI be?
98%
Ranked first, ahead of hosted Claude Opus 5
Local Qwen 3.8 27B matched or beat the field in 236 of 240 blind judgments across 80 synthetic legal matters, in CloseVector’s four-model comparison. Published August 15, 2026.
30 wins · 206 functional ties · 4 losses
WinsFunctional tiesLosses
Closed-record legal reasoning. Not a measurement of document-search recall.
Inspect the results and test recordsA benchmark is one test. Your IT and security reviewers will ask about sources, unread files and deployment. The buyer brief lists their questions.
Read the buyer briefHow the comparison was scoredCompany and principal
Built for work where getting it wrong has consequences.
CloseVector combines its own information-discovery platform with custom private-AI implementation. Founder Dean Hoffman brings experience building trading, risk-management and software systems to work where the records, the assumptions and the consequences all matter.
That background informs a simple discipline: question the result, keep the source, and make the work useful to the person responsible for the decision.
About CloseVector
Before anyone pays the $186,400.
equipment order approval
Purchase order 7316
- Order placed
- June 12
- Amount
- $186,400
June 12Order placedJune 18Approval cited
6 days apart
The question to resolveWhat authorized the June 12 order?
Search record savedIllustrative interface. Invented records.
Equipment order review: the six-day gap
Find the facts that could change your position while you can still act on them.
Prepare the next question before making the next commitment.
Walk into the review knowing what you can point to and what still needs explaining.
CloseVector helps you compare related records, with links back to the documents behind a finding.
Explore the workUse the information that matters
Put your most valuable information to work. Keep it under your control.
Search sensitive records and explore confidential questions without sending the files, prompts or analysis to an outside AI processor.
Focus on what you need to discover, rather than worrying about what the inquiry itself reveals to an outside provider.
The private system runs locally. A fully offline configuration is available when the work calls for it; access, updates and approved exports are part of the deployment design.
Hosted AI service
CloseVector
Ask of any AI service: where is the answer generated, and what leaves your environment?
Illustrative local configuration. Access, updates, backups and transfers are agreed for the deployment.
A colleague asks where the AI reviewed the confidential material. You can point to the system your organization controls, and explain what was allowed to leave.
Equipment order review: the supplier’s email
Read the whole promise before anyone acts on it.
Move from a finding to its source and context, so everyone works from the same words.
Your expertise goes into deciding what the record means, not retracing where the answer came from.
Source-linked findings and retained search records keep the path back to the material within reach.
See how the record stays connectedIllustrative example. Invented records.
The passage
“We will cover replacement costs…”
Read the full paragraph
The full paragraph
“We will cover replacement costs for the two evaluation units only. The production order is excluded.”
Two evaluation units. Not the production order.
When the team asks what is covered, you have the full passage ready.
Put private AI to work
Same private system. Different questions.
Illustrative examples. Invented records.
Incident · May 14AprilSuppliedMayMissingJuneSupplied
Legal & e-discovery
What could change your case?
Know what strengthens the case, what threatens it, and what to ask for next.
Explore the application$64,800+$7,200=$72,000LedgerAdjustmentReportedEmail: invoice hasn’t come in
Financial records
The numbers say one thing. The emails say another.
Trace the discrepancy to supporting records before signing off.
Explore the applicationBefore 09:4009:05Warning sent09:40Work began10:20First recorded acknowledgment
Internal investigations
Who knew what, and when?
Compare accounts with dated records and identify the next interview question.
Explore the applicationAsked: above 80°CEarlier seal trial
leakage at 85°C
leakage at 85°C
Confidential research
Your records are confidential. So are your questions.
Find earlier work that can inform the next research step, privately.
Explore the applicationA useful first step
Start with the question you need answered.
- 01
Tell us the question.
A few non-confidential sentences: the decision, the records, the control you need.
- 02
Get a written proposal.
What the system will do, what is included, what you will need to run it, and the price.
- 03
We set it up and check it.
Configured for your records and checked against agreed examples, with the results recorded.
- 04
Your team takes over.
Installed on hardware you own, with training, documentation and agreed support.
Selected research
A useful way to think about AI before you buy.
Evaluation
A better fit is not always a better answer.
Why out-of-sample tests matter, with interactive examples.
Read Curve Fit
Practical control
Know which exposure a control changes.
A visual guide to sensitive files, network boundaries and the work that remains.
Explore the guide
Operating discipline
Know what changed between answers.
Versions, custody, network access and upstream dependencies.
Read Live vs. PinnedStart a conversation
What would you like to be able to find out?
A few non-confidential sentences about the records and what makes the work difficult are enough to tell whether CloseVector is a fit.
Please do not send confidential files, client details, passwords or privileged material.
What happens next
We review your inquiry and reply by email. If there is a fit, we arrange a conversation about the first useful step.
contact@closevector.aiLocal preview: this form will not send an inquiry.