> This is the markdown version of https://www.maniac.ai/. Visit the full page for interactive content. recursive data analyst · built on your ontology # Ask the messy question. Get the deterministic answer. Arendil decomposes business questions into hundreds of grounded sub-queries across your warehouse, docs, tickets, and CRM, then reconstructs one cited answer. No SQL. No schema cleanup. No agent demo-ware. arendil · analyst · live listening question question 01question 02question 03question 04 real examples from enterprise users · cycling every few seconds [Ask a question of your data](https://app.maniac.ai/auth/register)[See how the ontology learns](/book-demo) soc 2 type ii deploys in your vpc bring your own model no schema cleanup \[ The moat is in the mess \] ## An analyst that reasons recursively. On data that was never clean. Arendil sits between your business questions and the ugliest schema in your warehouse, and makes the answer deterministic. 01 1,000× Sub-queries per question Resolved recursively in one user minute. 02 0 Schemas to pre-clean The ontology forms from your mess, not against it. 03 100 + Source systems unified Warehouse, SaaS, docs, tickets, code, comms. 04 15 min Average time from signup to first answer Plug in one warehouse and one system of record. \[ The shift under messy data \] ## Most AI flattens. Arendil digs in. Clean data is a commodity. Messy data is where every enterprise actually lives. Recursion is how you reason across it. FLAT AGENT Ask once. Answer once. Pray the columns line up. - One prompt → one answer → one table - Breaks on joins, nulls, misspelled columns - Gets worse as your data grows - Hallucinated insight, no lineage - You pre-clean the schema RECURSIVE ANALYST Every question becomes a tree. Every leaf is a fact. - Decomposed tree of sub-queries across every system - Ontology absorbs the mess before the LLM sees it - RL fine-tunes on every correction - Cited lineage, row-level provenance - The ontology forms itself \[ The ontology layer \] ## Your business has an ontology. You just haven’t written it down. Column names lie. Dashboards contradict. Definitions drift across teams. Arendil learns the true ontology from how your team actually queries, and makes it the contract every agent, app, and analyst shares. Learned, not configured Our RL loop catches the joins your docs forgot. Edges appear in amber the first time the system figures them out. Deterministic under messy input The same question resolves to the same query, every run. Reviewers can trust the number without re-reading the trace. Shared by every agent One source of truth, queried a thousand different ways. The ontology is the contract across the org. amber = the ontology figured it out· hover any field to see its sources ontology · customer concept · weaving sources 4edges learned 0 / 20 salesforce - Idstring - Namestring - BillingAddressaddress - OwnerIdfk - StageName\_\_cenum stripe - id - name - address - metadata.owner - status netsuite - internalId - companyName - billAddress1 - salesRepInternalId - entityStatus hubspot - hs\_object\_id - name - address - hubspot\_owner\_id - lifecyclestage concept customer 5 fields - idstring from 4 - namestring from 4 - addressaddress from 4 - owner\_idfk from 4 - lifecycle\_stageenum from 4 SELECT \* FROM ontology.customer Four sources, one concept. Every thread in amber is a mapping the ontology learned, not a join you had to configure. learnedsourceunified \[ Watch the analyst think \] ## Every answer is a tree you can inspect, replay, and challenge. One real business question. Decomposed into sub-queries. Each leaf grounded in a source system. A final answer with 14 citations. arendil · recursive query \[ Q3 pipeline investigation \] depth3leaves8sources6cites14 replay 01ask 02decompose 03resolve ask Why did enterprise pipeline slip in Q3? sub-query01 Stage transitions sub-query02 Call engagement sub-query03 Plan realism sub-query04 ICP segments \[01\]salesforce Stage B → C conversion −18% WoW \[02\]salesforce Stage C → Won conversion stable \[03\]gong Positive-sentiment calls −22% \[04\]zoom Average meeting length −6 min \[05\]confluence Assumed rep capacity +12% \[06\]snowflake Actual rep capacity +3% \[07\]snowflake Fintech usage −9% \[08\]snowflake Retail usage +2% answer · grounded cites \[01\] \[03\] \[05\] \[07\] · 14 total The Q3 slip concentrates at Stage B→C (−18%), where call sentiment collapsed (−22%) against a plan that assumed +12% rep capacity versus the +3% we actually shipped. Fintech ICP lagged (−9%) while retail held. inspect treechallenge stepexport sqlreplay \[ Every system becomes substrate \] ## Arendil learns the ontology that ties them. You don’t wire the joins. They emerge from how your team already queries. Every new system adds substrate, not setup. amber · learned relationships compound daily Warehouses source of truth SnowflakeBigQueryDatabricksRedshiftPostgres Systems of record transactional ground truth SalesforceNetSuiteWorkdaySAPHubSpot Tickets + knowledge narrative ground truth JiraServiceNowConfluenceNotionLinear Communication informal signal SlackTeamsGongGmailZoom Code + docs definitions drift here GitHubDriveSharePointDropboxBox \[ Why this compounds \] ## The messier your data, the stronger Arendil becomes. Recursive reasoning plus RL-trained ontology means accuracy goes up on the schemas your consultants failed on. day 1 Connect. Point Arendil at your warehouse and two systems of record. Mapping begins in the background. You don’t clean a single table. ontology coverage 15% resolved week 1 Ontology forms. ~70% of business questions resolve without human schema work. Learned edges light up in amber across the graph. Analysts stop being translators. ontology coverage 70% resolved month 1 RL compounds. Every correction trains the recursive analyst. Accuracy climbs above human baseline on your hardest schemas — and stays there. ontology coverage 100% resolved \[ Epistemic governance \] ## Audit the reasoning. Not just the output. Every run is a tree you can replay. Every correction becomes a learned edge. Every answer carries its provenance so reviewers can trust the number without re-reading the query. Replay the tree Walk every sub-query, see the exact row that produced each claim. Recompute with context When an edge is corrected, all affected answers re-resolve against the new ontology. Learn from challenges Human disagreement becomes training signal, not a bug report. Row-level permissions What the analyst can read is constrained by the same SSO + roles your BI tool uses. reasoning\_trace · today · 14:00–15:00 LIVE 1. 14:22:06query sarah.chen@ · cfo office Q: "Why did enterprise pipeline slip in Q3?" decomposed · 14 sub-queries 2. 14:22:09resolve analyst · recursive Walked 6 sources. Produced 14 facts with citations. grounded 3. 14:22:41correction marcus.oyelowo@ · rev ops Flagged: "Opp.StageHistory feeds PipelineView, not dbt\_pipeline." edge learned · ontology updated 4. 14:23:18recompute analyst · recursive Re-ran question against updated ontology. +2% closer to CFO baseline Every answer, every correction, every relearned edge. Stored, replayable, exportable. REPLAYEXPORTROLLBACK \[ Who it’s for \] ## Loved by the business. Trusted by IT and security. Give every team a faster way to get answers from messy data without bypassing the controls enterprise buyers need. Data + analytics leaders Answer the executive question before the dashboard loads. Reclaim your analysts from "can you pull..." tickets. Rev ops + finance Reconcile figures across Stripe, NetSuite, and Salesforce without a week of CSV surgery. IT + platform One governed surface for the analyst agent. One ontology. One auditable trail. Every team gets it. \[ Enterprise-ready by default \] ## The controls your procurement, security, and legal teams will ask for. Built to pass review at financial institutions, healthcare systems, and global services firms. Same controls whether we’re running in our cloud or inside yours. SSO, SCIM, SOC 2 Okta, Entra, SAML 2.0, OIDC, row-level access inherited from your warehouse. Bring any model OpenAI, Anthropic, open-source, or your own endpoint. We never lock the reasoning layer. Deploy in your VPC Managed cloud, single-tenant VPC, or on-prem. Your warehouse credentials never leave your boundary. Row-level lineage Every claim traces to the exact row. Every corrected edge is stored and replayable. Reversible by default Pause, rollback, or quarantine an answer path without taking the system down. Enterprise assurance Security pack, red-team reports, and procurement package available on request. \[ Next step \] ## Ask your first messy question. Plug in one warehouse and one system of record. See the ontology form in 15 minutes. Get your first cited answer before your next stand-up. [Start with your data](https://app.maniac.ai/auth/register)[Book a technical walkthrough](/book-demo)[Talk to our enterprise team](/contact) accuracy compounds after launchSOC 2 · VPC · BYO modelrecursion you can audit --- *Arendil, High throughput background agents. Opus-quality outputs at 1/50 of the cost. Learn more at [maniac.ai](https://www.maniac.ai).*