APEYRONAI
← Apeyron AI

Apeyron × HP Italy · Applied research partnership

Small local models. Engineered to be trustworthy.

Apeyron is an official HP Italy supplier. We build the software and teaching method that let non-programmers compose, test and understand bounded agent strategies on compatible local hardware.

HP Academy

The research thesis

The model is only one component. The harness is the product.

A small model becomes useful for a narrow task when the surrounding system knows how to retrieve evidence, separate specialist responsibilities, verify an answer, stop unsafe routes and fail honestly.

The collaboration turns those ideas into an executable visual language—so a teacher, student or evaluator can see what each step owns and why the activity continued, retried or stopped.

Workflow Lab

A strategy you can inspect before you trust it.

InputStudent requestUserMessage
GuardScope & safetyClassification
RetrieveApproved knowledgeEvidenceSet + provenance
GenerateLocal specialistDraftAnswer
VerifyEvidence checkApprove · repair · stop

Every connection has one documented meaning. A derived read-only graph helps explain the route; the canonical recipe remains typed and executable.

Engineering foundations

Reliability is made of visible decisions.

01

Typed responsibility

Models, classifiers, retrieval, deterministic tools and human decisions produce explicit artifacts instead of sharing one opaque context.

02

Bounded control flow

Branches are labelled, retries have a maximum attempt count, and every exhausted path has an authored safe response.

03

Evidence lineage

The final answer can cite only the evidence it actually consumed. Retrieval results keep their source and producing query.

04

Private observability

Teacher-safe traces retain routes, timings and call counts—not credentials, hidden reasoning, complete private chunks or student conversations.

Evaluation Lab

Same model. Same settings. A better question: what did the system change?

The laboratory compares a direct baseline with a reviewed workflow using the exact same local model snapshot and settings. A changed identity invalidates the comparison instead of being presented as equivalent evidence.

Deterministic route, policy, evidence and bounded-output facts remain separate from optional model-scored rubrics. Latency, model calls, Tool use and retrieval calls stay visible beside quality.

BaselineModel + instructionsone direct path
versus
Reviewed workflowModel + bounded strategyguards · evidence · verification

Local-first boundary

Private by default does not mean invisible engineering.

Inference stays localunless the creator explicitly adds a network capability.

Hardware is measuredinstead of assuming that an accelerator label guarantees model fit.

Logs stay boundedand exclude prompts, documents, credentials and private reasoning by design.

Failure closes safelywith an authored response when evidence, validation or budget is insufficient.

Partner with the lab

Have a difficult local-AI problem worth solving together?

We bring agent engineering, product software and education. We look for partners who bring a real domain, hardware or scientific problem—and want to share responsibility for the result.

Propose a collaborationRead the Italian story on Apeyron.it →