Evodira

EVODIRA

Governed distributed AI

Governed institutional memory

The files stay. The knowledge travels.

Evodira turns what your machines already hold into memory you can question. Every claim carries the quote it stands on, and a person decides before anything is trusted.

Become a design partnerSee how a claim is made

Early. We are validating this now with a small number of partners.

01 · Control

One control plane.

Identity, policy, fleet, releases and the audit trail live in one place. Your data does not.

02 · Sites

Everywhere your infrastructure lives.

A site connects through an optional fog gateway, or its machines speak to the plane directly. The middle layer is a choice, not a toll.

03 · Enrol

Give every machine an identity.

One command. The grant is one time, the key pair is generated on the device, and the private key never travels.

evodira-local-agent · the grant arrives on stdin, never argv

▊

During the exchange the device generates its Ed25519 key pair, reports hardware facts, pins the plane’s signing keys, and stores its credential in the OS keyring.

04 · Discover

Know what every machine can run.

Hardware facts only: cores, memory, accelerator, runtimes, and the names of installed models. Never file contents, never weights.

capability report · hardware facts only

node
gpu-nbo-01 · ONLINE
arch
arm64 · 10 cores
memory
16 GB installed · unified
gpu
metal · Apple Silicon
bandwidth
not measured · unknown, not slow
runtimes
ollama · llama_cpp
models
lfm2.5 1.2b instruct · lfm2.5 350m
agent
reports its version on every heartbeat

05 · Place

The workload goes where it belongs.

Every candidate that does not qualify is refused with the class of reason why, and the refusal is recorded.

BuiltBuilt and tested. Not yet wired into the running product.

placement · every refusal carries its reason class

pi-ldn-07refusedresource · memory_headroom
gpu-ldn-01refusedpolicy · outside the data boundary
ws-fac-11refusedevidence · unproven on this machine
gpu-nbo-01selectedmeasured, fits, inside the boundary

06 · Execute

Move computation to data.

A workload reaches one of three distances: the device, the site, or your own cloud. There is no public tier.

07 · Know

The file stays. The claim travels.

Redaction happens before any model sees the content. Only a typed claim leaves, with its quote, and a person approves it before it becomes memory.

context pack · the whole of what leaves the device

claim"Line 4 halts when chamber pressure passes 2.1 bar."
claim_type"process_rule"
supporting_local_refs[0].file_name"SHUTDOWN-RUNBOOK.md"
supporting_local_refs[0].page_number7
supporting_local_refs[0].heading_path"Shutdown > Overpressure"
supporting_local_refs[0].quote_hash"9f2c4a…"
confidence0.90
sensitivity"internal"
requires_reviewtrue

what the envelope has no field for

the documentthe page imagethe unredacted quotethe file path

08 · Operate

Deploy once. Operate the fleet.

Desired state is a counter every device converges on. Releases move through cohorts an operator promotes, behind gates that can refuse.

desired state · a counter every device converges on

desired rev 41converged 183 / 184behind 1silent 0

rollout · frozen cohorts, promoted by an operator

stage 1 · pilot machinescompleted 12 / 12
stage 2 · the wider fleethealth_check 138 / 240
gatesmax failure 5% · zero rollbacks · observed 30m

09 · Survive

Keep running when connectivity does not.

The fog gateway stores and forwards with signed hops. Work continues on site; nothing queued is lost while the WAN is.

fog gateway · store and forward, signed hops

wan
unavailable
outbound_queue
17 packs pending relay
policy_cache
served locally · age 240s
revocation_cache
enforced offline
device keys
signatures verified on site
on reconnect
drain until acknowledged, oldest first
09Architecture

Three planes. Only one of them ever holds your data.

Control plane

yours or ours
  • Identity and tenancy
  • Policy
  • Fleet and capability
  • Audit

Policy, metadata and receipts. Never source data.

Site node

optional
  • Shared compute
  • Local scheduling
  • Policy cache
  • Works offline

Only what the site boundary permits.

Device

where the data is
  • Direct source access
  • Local models
  • Extraction and search
  • Runs offline

The raw data. This plane, and no other.

Use the hardware you already have

  • Apple siliconMeasured
  • x86 servers and laptopsBuilt
  • ARM boardsDesigned
  • GPU workstationsDesigned
  • Industrial gatewaysNext
  • Your own cloudDesigned

Labels describe this repository today. Anything marked Designed or Next has a seam cut for it and no implementation behind it yet.

10Knowledge continuity

What your machines already hold becomes memory you can question.

This is the workload Evodira was built for. The runbooks, tickets and recordings that carry how your organisation works stay on their machines. What travels is claims: one statement each, typed, carrying the redacted quote and its source.

sourcesredactionlocal extractionevidenceclaimshuman verificationgoverned memorygrounded answer

Extraction runs on the device. Sensitive spans are redacted before any model sees them.

Every claim cites the quote it stands on, and travels with its provenance.

Nothing is trusted without a person deciding. The answer abstains when the proof is thin.

When the person who knew leaves, what they knew stays. When an auditor asks where an answer came from, the citation is already attached to it.

11Your plane

The address your fleet answers to is yours to publish.

A control plane is resolved up a three rung ladder: the fleet's own override, then the address your organisation publishes, then the platform default. HTTPS at every rung.

control.acme.ai

A fleet override. The first rung: one deployment profile points its devices at an address you run.

acme.evodira.com

A managed name. Claimed in the console, proved by the claim, published as the organisation's own sign in.

profile overridepublished planeplatform defaulthttps at every rung

Publishing a plane bumps the desired state of every fleet in the organisation, so devices re fetch and follow. Withdrawing it keeps who and when on the record.

12Models

Models are artifacts with evidence.

Evodira qualifies a model on the machine that will serve it, and keeps the receipts. Small Liquid foundation models are the working set, served through Ollama and llama.cpp.

An extraction rung

small enough for laptops

Turns documents into typed claims on the device that holds them.

A reasoning rung

the working default

Plans and answers where the extraction already happened.

A mixture of experts rung

for the machines that can hold it

More capacity, activated sparsely, still on hardware you own.

Identity is proven, not asserted

Manifest hash, file header geometry, and the digest the daemon reports after serving all have to agree.

Built

Benchmarks live on the machine

Tokens per second, first token, p95, per device and per model, with a sample floor. A changed environment marks them stale, never deleted.

Measured

Inventory tells its sources apart

A model reported by a task receipt carries its digest. A model merely named in a capability report never overwrites one that was verified.

Built

13Proof

One decision this actually made.

Recorded on a real machine. Everything above this is a claim. This is a receipt.

The receipt

Job
Classify a document
Ran on
the device itself
Chose
The smallest option that was good enough
Took
157 ms
Data moved
none
Considered
6 options, 5 refused

Why the others were refused

  • capability1

    Known to the platform, but not installed here.

  • evidence2

    Never measured for this job on this machine. Unmeasured is never treated as good enough.

  • resource1

    Would not fit in the memory this machine actually has.

  • selection1

    Good enough, and rejected anyway, because something smaller was also good enough.

Read the selection line again. Something bigger was good enough and lost anyway, because something smaller was good enough too. Most systems take the biggest thing that fits.

14Use cases

Different rooms. Same primitives.

Healthcare

Patient records stay on ward machines. Models move to the estate, and only approved outputs synchronise.

device boundary · local extraction · human review

Manufacturing

Models run beside production systems, and the fog gateway keeps the site analysing through a degraded WAN without losing a single pack.

site boundary · fog relay · store and forward

Enterprise knowledge

Private documents become governed organisational memory with citations, verification and provenance.

knowledge continuity · audit trail

Remote operations

Constrained devices keep working when the link drops, and reconcile when it returns.

cached policy · signed hops · offline queue

15Security

Private by architecture.

Boundaries
Device, site and your own cloud. Intersection only narrows, and an unknown value falls to the narrowest.
Signed hops
Devices sign every request with their own Ed25519 key. A fog gateway adds its signature and never replaces the device's.
Audit
A recorded trail behind its own read permission, one row per action, queryable per entity.
Keys
Generated on the device at enrolment. The private key never travels; the credential lives in the OS keyring.

Evodira is built for SOC 2 and GDPR from the architecture up, with data residency, redaction, audit trails, and human approval gates as first-class controls. The EU AI Act’s transparency and general-purpose obligations have applied since 2 August 2026, and we work to them as current law rather than as a date ahead of us. These statements describe our posture and design. They are not certifications we have completed. We will share the current state of our documentation, controls, and any attestations directly with you during a security review or pilot.

16Design partners

We are looking for a few design partners.

Organisations with real data on machines they already own, a reason it cannot be centralised, and a fleet varied enough that where things run is a genuine question.

A good fit

  • Data spread across devices, sites or a private cloud
  • A rule that says it cannot all be centralised
  • A mixed fleet, not one machine type
  • Willing to let us measure on your hardware

Not yet

  • You want a hosted assistant
  • Everything is already in one cloud with no constraint
  • You need production support this quarter
Become a design partnerSee the architecture

Evodira is not in production yet. The placement engine and the planner are built and tested, one workload is measured end to end, and a machine can now answer a fleet question. Nothing asks it one yet, because the console surface is still being written. If you need something to roll out to a fleet this quarter, we are not there.