Evidence collection, extraction, and review in one place.
Finance teams organize PBC evidence and run readiness checks. Audit teams review the same engagement with the extraction trail intact.
What the workspace does
Auditly keeps the request list, evidence intake, extraction results, readiness state, and auditor review tied to the same engagement record.
One engagement, scoped views
Client and auditor workspaces share the file while role boundaries keep each side's data scoped.
Evidence moves through a queue
Uploaded documents are scanned, parsed, classified, extracted, and surfaced when processing finishes.
Statement math is checked
Bank figures only move forward when beginning balance plus deposits minus withdrawals equals ending balance.
Preparation status stays current
The client can see what is complete, what is blocked, and which requests need attention before fieldwork.
Auditors work from the same record
The firm sees submitted evidence, extraction status, and review queues without rebuilding the client file by hand.
Every figure keeps its trail
Captured values keep method, model, confidence, and cost metadata for later review.
Seeded environment. The workspace uses synthetic companies and statements so it can be opened without private financial data.
Audit fieldwork starts with weeks of chasing paper.
The painful part of an audit isn't judgment — it's collecting, classifying, and tying out the client's evidence before any real work can begin. Auditly moves that work to the front, and automates the parts a machine should own.
Manual, back-and-forth, error-prone
PBC request lists over email. Statements re-keyed by hand. Tie-outs done in spreadsheets. The client shows up disorganized; the auditor spends the first weeks on logistics, not assurance.
Organized and half–tied-out on arrival
The client self-serves onboarding, evidence is extracted and reconciled on upload, and a readiness check tells them what's missing before fieldwork. The auditor opens a file that's already substantiated.
Four things worth building carefully.
Extraction is table stakes by now. The defensible parts are how the numbers get verified, how the AI stays honest, and how every artifact becomes durable evidence.
Deterministic-first extraction
Bank-statement specialists pull structured fields and reconcile them (tie-outs) before trusting them — most documents need no LLM at all. When one does, the call is metered and hard-capped, with a free local model as automatic fallback so the demo never breaks or overspends.
$0 on the deterministic pathGrounded copilot, allow-listed context
One reusable spine powers copilots on four surfaces. The model only ever sees an allow-listed grounding bundle and defers to the deterministic engine — it surfaces figures the engine validated, not ones it made up, and can't reach the auditor's internals. A safety-eval suite checks for leakage and fabrication.
no-leak / no-fabrication evalsEvidence substrate + signed audit trail
Every capture — extraction, verification, correction, metric — emits to an append-only log tagged with method, model, and consent. Control-test records are Ed25519-signed per tenant, so a modified or forged historical record fails signature verification — genuine tamper-evidence, not just an append-only claim.
append-only · Ed25519-signed recordsTwo-sided and multi-tenant
The same platform serves the audit firm and the client from distinct, role-aware workspaces. Isolation is enforced at the database with Postgres row-level security — a cross-tenant read returns 404, not 403.
RLS on every tenant tableHow a document becomes trusted evidence.
The interesting part isn't the box diagram — it's the trust boundary: an uploaded file crosses through deterministic validation and a reconciliation gate before it's ever written as evidence, and the LLM sits after the gate, not in front of it.
It reads the statement, then checks the math.
Most fields come from deterministic parsers. The model is only a fallback, and the result still has to tie out.
Real public formats for engine validation. Synthetic companies and statements in the seeded workspace.
Extraction performance is measured.
The engine is tested against labeled public bank-statement fixtures, with cost and latency tracked per run.
Product screens
A connected audit file from client preparation through auditor review.
Enter the workspace
Use the seeded client or firm account to move through the same engagement.
Prepare, extract, self-assess
Author, send, review
Working implementation
Auditly runs as a seeded product environment with a client workspace, auditor portal, background extraction jobs, AI assistance, and deployment infrastructure.
Project links
Code, evaluation notes, and contact details.