Sovereign · On-Premise · Industrial AI

Every source, every decision,
defensible.

hyperion-workbench is a sovereign, on-premise agentic AI workbench for confidential industrial work. Open-weight multimodal models, a local knowledge base, and bounded tools turn scanned documents into evidence-backed deliverables — with human decision authority intact.

Run the demo → The problem ↓

0

bytes of source data leave the boundary

0

recorded stages, capture → archive

0

consequential decisions confirmed by a human

Design invariants from the project requirements — not measured benchmark claims.

Confidential Industrial Operations

Confidential work is stuck between two barriers

The confidentiality barrier

Inspection reports, engineering calculations, drawings, SOPs and internal correspondence can't leave the organization — so conventional cloud AI is off the table.

The productivity barrier

Purely manual knowledge work — structuring reports, cross-referencing manuals, verifying claims — consumes engineer hours and loses decisions in the process.

Our operating model

source evidence→AI analysis→explicit inference→human decision

Not a chatbot. The system reduces repetitive knowledge-work effort without transferring accountability for consequential industrial decisions to the AI.

The engine we forked

We didn't build the harness from zero — we forked the best one

hyperion-workbench runs on our fork of DeepSeek Harness, the open-source all-plugin agent runtime. We chose it because the harness — not the model — decides how much of a model's capability becomes completed, reliable work. DeepSeek's own published benchmarks, all run through Harness minimal mode with fixed sampling and maximal reasoning effort:

BenchmarkV4-Pro GAV4-Flash
Terminal-Bench 2.187.982.7
Cybergym83.376.7
Toolathlon (verified)74.170.3
DSBench-FullStack71.168.7
DSBench-Hard67.259.6
DeepSWE62.754.4
NL2Repo61.554.2

Sources: V4-Pro GA (13 Aug 2026) and V4-Flash β (31 Jul 2026) — DeepSeek API Change Log, both run via DeepSeek Harness minimal mode (max effort, temp 1.0); DSBench-FullStack/Hard are DeepSeek's internal evaluation sets. Harness-swing comparison: independent rerun by Harrison Kinsley · Terminal-Bench context: tbench.ai 2.1 leaderboard.

Same model weights, different harness: an independent rerun of V4-Flash on a third-party harness swung Terminal-Bench by twenty tasks. That swing is exactly why our workbench fork is the investment: every capability — models, tools, skills, sessions, sandboxes, scheduling, even the UI — is a plugin we recompose for industrial workflows.

The nine-stage evidence path

Every run lands in one traceable log

hyperion-workbench — task trace / inspection-7291
captureinspection-report-scan.pdf · 14 pages · local store
understandOCR + vision · 6 tables, 2 drawings identified
groundSOP-4.2 r7 eff.2026-03 · pp.31–34 evidence linked
analyze3 findings · 1 conflict flagged · 2 high-risk claims
verifyengineer confirmed 2/2 · 1 correction applied
deliverapproval-note.docx generated · trace attached
reviewmanager approved · claim-level decision recorded
Source facts, AI inference and human decisions stay labeled and separate
Version and authority metadata surfaced on every source
Manager approves claims, not raw noise

P0 capabilities

Confidential in, evidence-backed out

Scanned PDFs, photographs, handwritten notes and drawings enter through a controlled local location. Open-weight models and bounded tools do the heavy work. Nothing leaves the boundary.

Multimodal capture

Local OCR and vision turn scans, photos and handwritten notes into inspectable text, tables and regions.

Grounded retrieval

A local knowledge base with version, authority and effective-date metadata. Every claim links to document, page and section.

Model router

Open-weight models registered with capability, modality and compute profile. Tasks route to the smallest model that can do the job — locally.

Sandboxed execution

Calculations and code run in resource-limited sandboxes with no unrestricted network — every invocation recorded.

Risk-based verification

Safety, equipment and numeric claims require human confirmation regardless of model confidence — grouped, not one dialog per claim.

Real deliverables

DOCX approval notes, XLSX outputs and working code — generated with their full evidence trail attached.

Sovereignty monitor

Visible proof, not promises

The PS asks for demonstrable isolation. The workbench exposes live execution state, controlled tool calls, and an event trail that shows sensitive tasks have no uncontrolled external path. Our sandbox plugins follow the same bet NVIDIA made with OpenShell — out-of-process policy enforcement, deny-by-default permissions, metered resources, and a full audit trail of every allow/deny decision.

Status labels describe deployment policy. The environment is validated before any air-gap claim is made.

SOV-01 / RUNTIME POLICY ENFORCED
SESSION — UPTIME 00:00 POLICY REV r7
External network
No unauthorized path
Model inference
On-premise · open-weight
Knowledge
Local controlled corpus
Sandbox
Active · resource-limited
Tool gateway
Allowlist · 12 tools
Audit trace
Recording task events

GATEWAY DECISIONS

    0 uncontrolled egress routes —

    Who the workbench serves

    Control is distributed deliberately

    Primary user

    Industrial Engineer

    Full workspace: create tasks, run analysis, verify grouped exceptions, generate deliverables. Evidence-first findings, never opaque conclusions.

    workspace · sources · artifacts

    Secondary user

    Department Manager

    Decision-focused review: summary, key findings, high-risk claims, evidence and engineer verification status — Approve, Reject or Request Changes per claim.

    evidence · flags · decisions

    Governance

    Cybersecurity Officer

    Live sovereignty proof: network isolation, least privilege, policy enforcement and the complete audit trail — without needing the workflow UI.

    isolation · policy · audit

    Governance

    IT / Knowledge Manager

    Read-only status: model registry state, corpus freshness and version/authority metadata, GPU resource profile — the platform stays governable.

    registry · corpus · metadata

    Clone the fork.
    Run it on the floor.

    The whole workbench is source-available — engine fork, plugins and all.

    Live on this machine

    The workbench is running — open the product directly.

    http://127.0.0.1:3080

    Clone the fork

    The workbench fork — HYPERION brand baked in via profile.

    $ git clone https://github.com/Narendarcodes/hyperion-workbench.git
$ cd hyperion-workbench && pnpm i && pnpm dsh --profile web
    Open the workbench →