← Novan Labs

Dialect 01 · For private equity & real assets

Document diligence at machine speed.

Your analysts spend hundreds of hours reconciling a data room, leases, rent rolls, T-12s, estoppels, PSAs, loan docs. We build the instrument that extracts every material term, cross-references them for contradiction, and red-flags the risk, each finding cited to the clause. It feeds your platform. The analyst stays the author.

Start a sprint → See the work
01 · The problem

The risk hides between the documents.

A single lease is easy. The exposure lives in the reconciliation, the rent roll that disagrees with the T-12, the estoppel that contradicts the lease, the co-tenancy clause that nobody flagged, the indemnity that was never capped. Doing that by hand across a full data room is slow, and the miss is silent until it's expensive.

The risk isn't in any one document. It's in the contradiction between three of them.
02 · The instrument

Extract. Cross-reference. Cite. Score.

Four moves, run on your documents, instrumented end to end. Not a chatbot over your data room, a graded pipeline with a number on the front.

Move 01Extract
Every material term, dates, dollars, parties, options, caps, escalations, pulled from each document into a structured, typed record. Scanned PDFs included.
Move 02Cross-reference
The pipeline compares records across documents and surfaces the contradictions a human would need to hold four files in their head to catch. Cross-document findings are the priority.
Move 03Cite or it doesn't ship
Every finding must cite a clause that exists in the input, document and reference. A claim with no citation is discarded by design, not surfaced and hoped over.
Move 04Score
Findings are severity-ranked with a human sign-off column. The analyst reviews, edits, and remains the author of record. The AI stays inside a reviewed envelope.
Or watch it happen, once.
lease_04.pdf · §3.1, §41.2scanning…

…Tenant shall pay Base Rent of $18,500 per month, increasing by 3.0% per annum on each anniversary of the Commencement Date…

rent_roll_q2.xlsx · row 7scanning…

…Suite 240 · current rent $17,250 / mo · escalation 2.0% · term through 2031-03-31…

estoppel_s240.pdf · ¶4scanning…

…Tenant certifies that rent is current, that no amendments exist except as attached, and that no defaults are outstanding as of the date hereof…

SEV-1 · cross-documentBase rent disagrees across documents. The lease says $18,500/mo; the rent roll says $17,250/mo, a $15.0K/yr gap in underwriting. Cited: lease_04 §41.2 · rent_roll row 7.
SEV-2 · cross-documentEscalation mismatch. 3.0%/yr in the lease vs. 2.0% in the rent roll, ≈$4.9K/yr of drift by year three. Cited: lease_04 §3.1 · rent_roll row 7.
status 2 findings · sign-off pendingfields cited 6run cost $0.041

A scripted demonstration on synthetic excerpts. The production pipeline runs on your documents, under NDA, with the meter on, and a finding without a citation never ships.

03 · Where it sits

It feeds your platform. It never replaces it.

Our instrument is a bottom-up document layer that hands structured, cited output to your system of record, the place your firm already trusts. We complement the pipeline your team built; we don't compete with it.

Input

Data room

Leases, rent rolls, T-12s, estoppels, PSAs, loan docs, as they arrive.

Novan-1

Extract

Material terms → structured, typed records. Scans included.

Novan-1

Cross-reference & flag

Contradictions surfaced, clause-cited, severity-ranked.

Yours

System of record

Structured JSON into your platform, your schemas.

Yours

Analysts decide

Human review, sign-off, authorship. The judgment stays with you.

Novan builds & hands youYours, untouched
You own everything with your fingerprints on it. We own the machine shop that made it.
04 · How we prove it

The scorecard is the deliverable.

We don't ask you to trust the model. We measure it, on your documents, against your people, and hand you the measurement.

Evaluated against your analyst baseline.

Ground truth is built with your experts on a held-out set of your documents. We score field-level accuracy against that baseline, hold it to an acceptance bar we set together on material extractions, and report the misses as findings, because the miss is exactly what the eval harness exists to catch. When the model is wrong, that's a data point you paid almost nothing to learn.

Accuracy scorecard vs. baseline Cost metered per deal file Latency stamped per run Pinned model versions Full audit log
05 · Governance

Table stakes, in writing, before any data moves.

Every engagement maps to the NIST AI Risk Management Framework, Govern, Map, Measure, Manage, and ships a governance one-pager your compliance team can file.

TenancyIsolated deployment
Your documents run in an isolated environment. No shared context, no co-mingling with other engagements.
TrainingNever on your data
Your data trains no model, ours or the provider's. Deletion with written attestation on request.
AuditabilityEvery decision logged
Inputs, outputs, model version, and cost recorded per run. Reproducible, inspectable, exportable.
DriftRe-evaluated on every bump
Models are pinned; the full eval suite re-runs before any version changes in production, and drift deltas are reported first.
06 · The offer

A 14-day sprint. The audition is free.

We run the instrument on a slice of your real work and hand you the evidence. You decide with a scorecard in front of you, not a pitch.

You send
A representative set of documents under NDA. One deal is enough.
You get back
A findings memo, an accuracy scorecard vs. your analyst baseline, and a governance one-pager, inside two weeks.
You keep
The eval harness and the artifacts, whether or not we continue. If the scorecard says you don't need us, that's a cheap answer to have bought.
Then
If it clears your bar, we scope the production build, fixed fee, per tool, with a start date in the room.
Start

Bring us one deal's documents.

Tell us where diligence stalls today. We'll show you the instrument on your own data, under NDA, metered, and scored.

Start a sprint → hello@novanlabs.ai