Agentic solutions
Agents that use tools, query internal data and work together, with every action recorded and under policies you define.
onpremIA Labs: a lab building agentic and intelligent solutions for local AI.
We design, evaluate and deploy artificial intelligence that runs on your own infrastructure, and release as open source the tools to audit it, operate it and learn with it.
We build AI that is gobernable
What we do in the lab
We research, prototype and take to production intelligent systems designed to run outside the cloud: on your own servers, GPU workstations and edge devices. What we build for ourselves, we publish for the community.
Local and edge AI
Language, vision and speech models, quantised and deployed on your hardware, even with no internet connection.
Audit and certification
We measure how models, agents and whole systems behave, and turn it into signed certificates anyone can verify.
Talent and training
Platforms that bring training closer to real work, because technology is only sovereign if there are people able to run it.
- iLocal firstIf it can run on your infrastructure, it runs there.
- iiMeasurableNothing is claimed without a measurement a third party can repeat.
- iiiOpenTools are released under a free licence.
- ivHonestWhat could not be measured is declared, never hidden.
Our values
Five principles that become concrete engineering decisions, and that you can check in what we publish.
Governance
01An AI's decisions have an owner and a trail.
- Hash-chained, sealed records
- Policies with comply, violate or undecidable verdicts
- Doubtful actions held for human approval
- Thresholds are set by the operator, never by the tool
Sovereignty
02Your data, your models, your infrastructure.
- Works with no network connection (air-gapped)
- No mandatory dependency on the cloud
- Certificates verified with no server and no network
- Open source: no black boxes
Security
03Secure by design, minimal by default.
- Ed25519 signatures over canonical JSON (RFC 8785)
- Minimised evidence: fingerprints, never messages or secrets
- Datasets and auxiliary models pinned by hash
- Prompt-injection scenarios with canaries
Identity
04Every component knows what it is, and can prove it.
- Deployment fingerprint: model, quantisation, engine and hardware
- Identification of agents, models and tools
- Actor chain: which agent acted on whose behalf
- Issuers with DIDs and W3C verifiable credentials
Explainability
05Results you can explain and repeat.
- Results per category, never an overall grade
- Deterministic detection, no “LLM as a judge”
- What was not measured is declared; zero means “measured, and it was zero”
- Methodology and decisions published, with their cost
Lab products
Auditing what already runs, building and deploying what comes next, and training the people who will make it happen.
Audit
Live
OPAAS
Audits models, agents and whole systems that are already running, and issues signed certificates that verify offline.
7audit modes
Platform
In development
Evolve
A self-sufficient platform to adapt, orchestrate, evaluate and deploy multimodal AI on on-premise hardware and edge devices.
7AI modalities
Education
In development
NexoDual
A digital platform that connects academic talent with the world of work, centralising vocational training through dual programmes.
3actors on a single platform
No products in this category.
Live · Open source
OPAAS
On-Prem AI Audit Suite
Evaluates an AI deployment that is already running —without deploying or downloading the model— and turns the result into a signed digital certificate a third party can verify offline.
- Model performance, capability, consistency, safety and compliance.
- Agent auditability with instrumented ground truth, not an LLM judge.
- Whole-system audit: models, agents, tools and records.
- W3C credentials signed with Ed25519, and continuous change watch.
What OPAAS evaluates in a model
Performance
A- Tokens/s and TTFT
- Latency under concurrency
- VRAM, power and hardware fingerprint
Capability
B- Instruction following (IFEval)
- Answering from context (SQuAD 2.0)
- Abstaining when it does not know
Consistency
C- Determinism across repetitions
- Stability under concurrent load
- Robustness to malformed input and truncated context
Safety
D- Over-refusal (XSTest)
- Bias per social category (BBQ)
- Toxicity under provocation (RTP)
Compliance
E- Requirements cited to the clause
- Declared and verified, always apart
- What is unsupported is marked as such
And beyond the model: seven audit modes
benchCertifies a model in its real deployment.discoverInventories engines, agents and tools.smithScenarios with mock tools as ground truth.swarmDeclared against observed multi-agent flows.sentinelA proxy that observes and holds actions for a human.inquestPost-hoc audit from records.auditThe whole system, with a single certificate.
In development
Evolve
No-code · low-code · air-gapped · self-sufficient
A platform to adapt, orchestrate, evaluate and deploy multimodal artificial intelligence locally —vision, speech, SLMs, LLMs, agents, RAG and graphs— on on-premise hardware and edge devices, without relying on external services to run.
- Visual flow design for those who do not code; extensible with code for those who do.
- The full cycle inside your network: from tuning the model to deploying and evaluating it.
- From the GPU server to the edge device, with the same flow.
- Self-sufficient: no cloud licences and no external calls at run time.
Evolve is in active development. We will publish the repository once its first version is usable.
- 01Adaptfine-tuning, quantisation, distillation
- 02Orchestrateagents, tools, RAG and graphs
- 03Evaluatereproducible metrics before deploying
- 04Deployon-premise and edge, offline
VisionSpeechSLMsLLMsAgentsRAGGraphs
On-prem servers
GPU workstations
Edge devices
In development
NexoDual
Academic talent ⇄ world of work
A no-code digital platform that connects academic talent with the world of work and centralises vocational training through dual programmes, bringing schools, companies and students together from the agreement to the final assessment.
- Agreements and training plans in one place.
- Shared placement tracking and tutoring.
- No code: each school configures its own programmes.
NexoDual is in development. If you represent a school or a company interested in dual programmes, let's talk.
Pick an actor to see what it brings and what it gets.
A single link
Agreements, training plans, placement tracking and assessment, shared by all three parties without scattered paperwork.
Open source and community
A certificate only its issuer can check is not a certificate. We publish the code, the methodology and the decisions so anyone can audit the auditor, and we work alongside those driving local AI forward.
MITFree licence, commercial use included
W3C VCVerifiable credentials, an open standard
RFC 8785Everything signed goes through canonical JSON
0External services needed to verify
Collaborators
The main driver of the artificial intelligence ecosystem in Chile, promoting responsible innovation with national and Latin American impact.
Online store specialised in collectibles.
Cloud and CDN services provider. Collaborates on scalability and hybrid deployments of local models.
Sponsors & Partners
Shall we build on-premise AI together?
Evaluate your infrastructure with OPAAS, take part in the development of Evolve, or bring NexoDual to your school or company. Write to us and let's talk.