EigenForge AI Labs Talk to us
Layered geological strata of platforms with one glowing coral seam

Platforms

Two production platforms. One accountable partner.

We do not ask clients to assemble a stack. ZigmaData makes your data answerable; Kyros makes the agents that act on it accountable. The platforms are named here because you should know exactly what you are getting.

GET DATA INConnectComposeconnectors + patent · batch, CDC, streamMAKE IT USABLEContainConvergeCurateopen columnar · entity resolution · quality on arrivalMAKE IT GOVERNEDCatalogControlmeaning and lineage · policy at query timeMAKE IT ANSWERABLECognitiveConsumeontology retrieval · SQL, language, BITHREE NON-NEGOTIABLES✓Read-only at source✓Open storage formats throughout✓Ontology grounding, not a vector copy
Exhibit 01ZigmaData: nine components in four bands, and three things it will not compromise on.
01Already builtA large connector library across ERP, database,application, stream, file and object sources.02Or days, not a projectA new connector is routine work rather than anintegration programme or a change request.03Or the patented routeAny external program queried as though it were astandard table. Nothing is unreachable.Logic already built in source systems — database views, CDS views, APIs — is consumed rather than re-implemented.There is no transformation layer for you to maintain.
Exhibit 02Connectivity in three steps. The third is the patented route, for systems with no interface at all.
THE COMMON APPROACHChunk, embed, retrieve bysimilarity✕Parent–child relationships are lost✕Joins cannot be rebuilt from similarity✕The copy drifts from the source✕Fluent, confident, sometimes wrong✕Content leaves the boundary to be embeddedAN ONTOLOGY OVER LIVE DATAModel relationships, resolve,query✓Table relationships preserved, joins correct✓Answers generated from live records, never from a stale copy✓Every answer traces back to records✓Policy applied at query time, per user✓Runs against a local model — nothing leavesIs it answering from a copy of your data, or from your data?
Exhibit 03The distinction that decides whether an answer can be trusted — and the question to put to any vendor.
FIG. 1 RETRIEVAL IS NOT ONE THING — FIVE PATTERNS, GRADED HONESTLYPATTERNWHAT IT ISWHERE IT WINSWHERE IT FAILSUSE IT WHENNAIVE RAGchunks, embedded, retrieved bysimilarityfast to stand updrifts from source;relationships lostdocuments that stand aloneADVANCED RAGre-ranking, query rewriting, HyDEbetter recall on messycorporastill a copy, still ageingpolicy and handbook librariesMODULAR RAGswappable retrieve / rerank /generate stageseach stage measuredseparatelyintegration burden is yoursteams with evaluationdisciplineGRAPH RAGentities and relations retrieved asstructuremulti-hop questions,lineageover-applied where a tablewould doquestions that span recordsAGENTIC RAGan agent plans retrieval as asequence of callshard, underspecifiedquestionscost and latency compoundresearch-grade questions,budgetedThe last two rows are where enterprise questions actually live — anything whose answer spans more than one record. That is theretrieval problem our research programme works on.
Exhibit 04Why we resolve questions against an ontology over live records rather than a vector copy — the five retrieval patterns, graded honestly.
EVERY FIGURE ARRIVES WITH ITS LINEAGE ATTACHEDthe ERPthe CRMthe documentsTHE GOVERNEDLAYERdefinitionsqualitypolicylineagethe answerEVERY MARK TRACES TO A SOURCEREAD-ONLY · NOTHING COPIED · RECORDED AS OF THE DAY
Exhibit 05Watch an answer assemble: sources flow in, the governed layer binds them, and what lands carries its citations.

The data platform · Patented · in production

ZigmaData

A unified real-time data platform that reads every system you already run — in place.

Delivered by EigenForge on the ZigmaData platform, which is patented and in production. EigenForge is a delivery and go-to-market partner with engineering capability on the platform.

Nine integrated components deployed as a single auto-scaling cluster. It reads your systems of record read-only — no migration, no write-back, no change to transaction processing. It governs what it retrieves: definitions agreed once, quality applied as data arrives, lineage on every figure, policy enforced when a query runs. And it answers: SQL for engineers, plain language for everyone else, your existing BI tools over the same layer, and agents that raise the exception without being asked.

Switch it off and everything carries on exactly as before. That single property is why it gets through security review when a migration programme would not.

What sets it apart

Ontology-grounded AI
Natural-language questions resolve against an ontology built over live data. Relationships between entities survive, joins stay correct, and every answer traces back to a record. Not a vector copy that drifts from its source.
Open by design
Open columnar formats throughout, so your data stays portable and readable by other tools. Lock-in stops being a procurement concern.
Deployment independence
Standard Kubernetes rather than a provider-specific managed service. One architecture serves a defence tender, a bank's risk committee and a cloud-first CIO.
AI inside the boundary
Inference runs on local models inside your environment. No data leaves, and no per-token cost.
Integrated, not assembled
Cataloguing, quality, entity resolution and integration are native components of one product instead of four separate procurements.
Protected innovation
The method for querying any external program as though it were a standard table is protected by a granted patent. Nothing in an estate is unreachable.

What is in it

Connect

Connectors, plus the patented program-as-table method

Compose

Batch, change-data capture and streaming

Contain

Open columnar storage, tiered

Converge

Entity resolution across systems

Curate

Quality rules applied as data arrives

Catalog

Meaning and lineage

Cognitive

Ontology-grounded retrieval

Consume

SQL, natural language and BI

Control

Row and column policy at query time, SSO, encryption, full audit

What it connects to

SystemHow it is readWhat you get
SAP · S/4HANA · ECCTables, CDS views, OData servicesExisting logic reused
Oracle · Dynamics · other ERPDirect, incremental or API-basedFull history accessible
Core banking · policy adminDatabase, file or API readRegulated data unified
CRM · HR · service platformsAPI and object readCustomer and workforce joined
Mainframe · legacy applicationsProgram-as-table (patented)No source excluded
Warehouses · lakes · databasesDirect query, no copyPrior investment reused
Streams · IoT · filesStreaming, object and file ingestOperational data in context

An extensive connector library already spans every major category. A new connector is typically days of routine work rather than an integration project. Where no connector is possible at all, the patented program-as-table route makes the source queryable anyway.

Channelsvoice, chat, web, analyst co-working, offline field clientAgent runtime & orchestrationapproval and interrupt as primitives, long-running workflowsPolicy engine & approvalsaction gates on the decision path, approval routingEvidence & attestationsealed record per action, artefact hashing, model provenanceMemory, graph & datagraph and vector memory, governed lake, hybrid search, temporalityConnectors & sandboxed executionopen protocols, container and micro-VM sandboxesACROSS EVERY LAYERIdentity and entitlement forpeople and agents alikeAny-model routing, withself-hostingOn premises, at the edge, orair-gappedResidency and redaction byjurisdictionEncrypted mesh between sitesPerimeter-native: the engine runs inside your boundary and the audit is signed at the edge.
Exhibit 06Kyros, powered by Kaman. Policy and evidence are the load-bearing layers; everything else is table stakes.
FIG. 1 A PRODUCTION AGENT IS A LOOP, NOT A MODELTHE MODELrented, swappablePOLICY & BUDGETper-agent spend counters and hard limits,enforced outside the modelGATESauto / approve / advise per action class,decided before anything is builtHARNESStools with budgets, memory that isinspected, a sandbox, an evaluation loopEVERY LAP OF THE LOOP LEAVES A SIGNED RECORD
Exhibit 07A production agent is a loop, not a model. Kyros is the harness: tools with budgets, gates on the decision path, a signed record of every lap.
An agentproposes an actionPolicy engineon the decision pathnot at the front door — scheduled jobsnever come through the front doorAUTORoutine, reversible actions within set limits. Fullyrecorded.a reminder · a classification · a file movementAPPROVEThe agent prepares the action and its evidence; anamed person approves before anything happens.most work with financial or operational consequenceADVISEThe agent informs; the person decides.the default wherever entitlement or liability is involvedAutonomy is granted by action class, not by agent — and agreed before anything is built.
Exhibit 08Autonomy granted by action class, enforced on the decision path — not at the front door.
Requestwhat was asked, by whomEvidencethe records it reliedonPolicy checkthe gate and the ruleApprovalwho, when, what theysawActionwhat was done, whereSealhash of this and thepreviousEach entry carries the hash of the one before it, so any deletion or reordering is detectable.Any decision can be explained from the data as it stood on the day — including which model ran, with which prompt.
Exhibit 09Six parts, hash-chained, and reconstructable as of the day the decision was made.

The agentic platform · In production · partner platform

Kyros (Powered by Kaman)

A perimeter-native platform for governed, agent-native operations.

Delivered by EigenForge on the Kaman platform, which is in production. EigenForge is a delivery and go-to-market partner with engineering capability on the platform.

The engine runs inside your boundary. A policy engine sits on the decision path of every agent action. A tamper-evident audit is signed at the edge. The point is that agents can do real work — reading, reconciling, preparing, routing and acting — while every action stays bounded by a named person's authority and remains reconstructable afterwards.

Agent programmes rarely stall on capability. They stall on three questions: where would data leave the boundary, what happens when an automated step affects a person, and can anyone reconstruct afterwards what the system did and why. This is built around those three.

What sets it apart

Policy on the decision path
Every action class carries a gate, enforced where the work is done instead of at the front door — because scheduled jobs and background processes never come through the front door.
Signed, reconstructable record
Each action is signed by the engine that performed it and can be reconstructed as of any past date, including which model ran with which prompt.
One control plane over a mixed estate
Most organisations will run agents from several vendors. Policy and audit apply to every agent, not only the ones we built.
Reach where others stop
Connectors over open protocols including IBM i green-screen, sandboxed execution, and governed screen automation as a bridge where no interface exists.
Sovereign by construction
On premises, at the edge, air-gapped or private cloud — the same product, configured differently, with any model interchangeable.
Acceptance is a captured run
Every user journey is a document, an executable test and its captured proof. Transitions are blocked without evidence.

Our governance platform

And Mnemos, underneath our own products.

The platform every product we build stands on — built once, so governance is never re-argued. It is in development and is not sold on its own.

Mnemos, the platform underneath

Put the hard questions to us

Architects and security reviewers get a direct answer, in writing, including where the honest answer is a limitation.

hello@eigenforgelabs.ai

Send opens your email client with the note already addressed to us — nothing is stored on this site, and the message goes from your own mailbox, so our reply lands in yours.

Keep reading