Platform

The Numeri platform

A multi-agent system built for professional work, where output must be verifiable before it can be relied on.

Why multi-agent

A single model handling an entire engagement fails on professional work. Context degrades across long documents, arithmetic is unreliable, and nothing independently checks the result. Specialized agents with narrow responsibilities perform better, and the system's behavior is inspectable at every stage. When something is wrong, you can see which stage got it wrong.

The six stages

The five agents already live in the product (Intake Agent, Bridge Agent, Cash Agent, Red-Flag Agent, Narrative Agent) are the financial due diligence instance of these same six stages, named in the table below, not a different or superseded system.

hands off to normalize

hands off to domain work

hands off to verify

hands off to exception

hands off to output

Stage	Agents	What happens
1. Intake	Intake Agent (data quality assessment)	Classification and extraction from documents as they actually arrive, every value tagged with the page and position it came from.
2. Normalize	Intake Agent (chart mapping, period alignment)	Extracted data mapped to consistent structure: chart of accounts, trial balance, period alignment, entity separation. The bulk of manual preparation time lives here.
3. Domain work	Bridge Agent (EBITDA adjustment, quality of revenue, historical trading, product and service line), Cash Agent (proof of cash, net debt, NWC, cash flow and capex), Narrative Agent (outturn and forecast review)	Service-line analysis configured to your firm's methodology during deployment. Generic defaults produce output reviewers do not trust, so there are none.
4. Verify	Recomputation Verifier computed, Transaction Support Verifier, Bridge Integrity Verifier	Every claim another agent produces is independently rechecked: figures recomputed from source, citations confirmed to exist and support the proposition, extractions compared against the original document. What fails verification is withheld and flagged with the reason. No agent verifies its own work.
5. Exception	Red-Flag Agent	Items requiring professional judgment route to a qualified person with the document, the ambiguity, and the alternatives attached.
6. Output	Narrative Agent, Bridge Builder, Databook Assembly, Excel Export	Deliverables in your formats: your templates, your indexing, your tickmark legend, Excel that preserves your structure.

How the platform stays reliable

Deterministic computation

Language models are unreliable at arithmetic and confident when wrong. All calculation runs in verified code. Models read, classify, and explain. Code computes. The two are compared, and disagreement is surfaced for review.

Traceability

Every value in every output links to its origin: document, page, and position, or the ledger row or export line it derives from. Any figure opens to its source in one step. An output that cannot be substantiated costs more to verify than to reproduce.

Review and audit trail

Every proposed item is reviewer-controlled: accept, reject, or modify, each action logged with reviewer, timestamp, and rationale. That log is your record for quality control, peer review, and the question that arrives eight months later.

The pipeline for your service line

The architecture is constant. The domain agents and review gates change by service line: a valuation pipeline routes every proposed adjustment to a human, while a CAS pipeline clears high-confidence categorizations and escalates the rest. Every service line on this site has a specified pipeline already, named agents, the verification stage, and exactly where exception routing fires, not a placeholder. What a pilot builds is that design running against your templates and your data. Each service page publishes its full pipeline: stages, agents, computed steps, and what your reviewers hold.

How tax runs

11 agents · 6 stages

1
Stage 1 · Intake
- K-1 Agent
  Classifies every K-1, 1099, and organizer document as it arrives.
- Bundle Agent
  Splits one combined scan into its individual documents.
2
Stage 2 · Extract and normalize
- Box Agent
  Extracts every box and code into one reviewable schedule.
- §199A Agent
  Parses §199A, §163(j), and the detail buried in footnotes.
- Basis Agent
  Rolls forward capital accounts, basis, and passive-loss positions.
3
Stage 3 · Domain work
- Apportionment Agent
  Computes state apportionment and K-2/K-3 mapping by jurisdiction.
- Safe Harbor Agent
  Computes extension and estimated-tax safe harbors, client by client.
4
Stage 4 · Verify
computed in code
- Footing Agent
  Recomputes every total independently of the agent that drafted it.
- Diagnostic Agent
  Flags every material movement against last year’s return.
No agent verifies its own work. This stage re-derives from source. What fails is withheld and flagged, never passed downstream.
Failed checks route to your reviewers
5
Stage 5 · Exception
- Referral Agent
  Routes ambiguous footnotes and unusual codes to a preparer, document attached.
Judgment items route to your reviewers
6
Stage 6 · Output
- Filing Agent
  Formats the reviewed schedule for import into your tax software.

Activity, across engagementsIllustrative

Book a workflow assessment

The systems we build against

Numeri does not require a pre-built integration to start. Our engineers work with whatever your firm already exports, Excel, CSV, PDF, or a direct pull where the software supports one. That is what a forward-deployed engineer does differently from a platform that only works with what it already connects to. Below is the software already common across each service line, so you know we are not starting from zero on yours.

Service line	Systems we build against
Tax	UltraTax CS, Lacerte, CCH Axcess Tax, Drake Tax, GoSystem Tax RS, SurePrep, Checkpoint and CCH AnswerConnect for research
Audit and assurance	CaseWare, CCH Axcess Engagement, TeamMate+, Confirmation.com, IDEA
Client accounting (CAS)	QuickBooks Online and Desktop, Xero, Sage Intacct, NetSuite, Bill.com, Gusto, ADP, Dext, Hubdoc
Advisory and CFO	Jirav, Fathom, LiveFlow, Mosaic, QuickBooks Online, Sage Intacct
Valuation and forensic	PitchBook, Capital IQ, DealStats, GF Data, Kroll Cost of Capital Navigator
Transaction advisory	Datasite, Intralinks, Firmex, Ansarada, Excel
Practice operations	Karbon, Canopy, CCH Axcess Practice, BQE Core, Practice Ignition

Don't see your software named? That's fine, tell us on the call. If your firm can export it, we can build against it.

Book a workflow assessment