Solutions · Kiosk

Voice-first kiosks with a per-answer audit trail.

MOS4 kiosk delivers on-device LLM, refuse-preferred RAG, four-layer prompt-injection defence, and multilingual voice — all offline by default. Every answer carries a full provenance record.

AI · kiosk vertical
≤ 280 MB LLM on-device RAM acceptance criterion: 4-bit quantised SmolLM2 360M on compute-class
≤ 15% WER at 70–75 dB acceptance criterion: voice in industrial noise floor
6 months audit retention per-answer provenance record, rolling window

Differentiator matrix

MOS4 kiosk versus alternative approaches.

Category labels per decision D-02 — no named competitors. Each row is an independently verifiable capability, not a marketing claim.

Kiosk approach comparison — MOS4 kiosk vs category alternatives
Capability MOS4 kiosk Cloud-native SaaS Hardware-OEM + AI partner Conversational agent platform
Offline-first answers yes no partial no
Documented RAG refuse-gate yes rare no no
Per-answer audit manifest yes no no no
4-layer prompt-injection defence yes partial partial partial
Single-vendor full stack yes no no no

Cloud-native SaaS

Always-on connectivity required

cloud dependent

  • Offline operation not available
  • Audit trail depends on cloud logging
  • RAG refuse gate rare or absent
  • Prompt-injection defence partial
  • Multi-vendor integration required

MOS4 kiosk

Offline-first by design

single-vendor full stack

  • LLM, RAG, and voice run on-device
  • Per-answer audit manifest on EventBus
  • Documented RAG refuse-gate with cosine threshold
  • Four-layer prompt-injection defence
  • Single vendor from camera to audit log

Evidence

Four pillars of the MOS4 kiosk platform.

Refuse-preferred answer

The RAG pipeline checks cosine similarity before generating a response. Top-1 chunk: ≥ 0.55. Top-3 mean: ≥ 0.45. Below threshold, the kiosk says "I do not have information on that." Acceptance criterion: ≥ 80% refuse-rate on out-of-corpus, ≤ 10% false-refuse on in-corpus.

See RAG platform →

Voice in noise

Whisper-tiny multilingual with mandatory STT vocabulary boost for domain-specific terms. Acceptance criterion: WER ≤ 15% at 70–75 dB — typical factory floor and lobby noise levels. Streaming TTS with first-audio target of approximately 200–300 ms.

Audit trail

Every answer publishes a provenance record on the EventBus: chunk IDs, document paths, model version, similarity scores, and refusal reason. Six-month rolling retention. Suitable as evidence input for EU AI Act §10 and §13.

See EU AI Act posture →

MCP tool gate

The model can only call tools explicitly listed in the MCP allow-list. Default list is minimal. Tool escalation requires operator configuration. Acceptance criterion: ≥ 95% deflection on 20-prompt red-team injections.

Scope boundaries

What MOS4 kiosk does not cover.

Visual PPE / CV detection

Computer vision and PPE detection are covered by the AI Vision platform, not the kiosk platform. Both can run on the same device.

See AI Vision →

Cloud orchestration

Heavy-duty cloud orchestration, multi-region routing, and managed LLM API calls are out of scope. The kiosk runs at the edge; cloud is opt-in, not the default.

Foundation-model training

MOS4 kiosk deploys pre-trained quantised models. Foundation-model training and fine-tuning workflows are out of scope. RAG and vocabulary boost cover domain adaptation.

Explore further

Related capabilities.

AI Language platform

The full on-device LLM, RAG, MCP, and voice platform that powers the kiosk vertical. Five pillars with acceptance criteria for each.

See AI Language →

Compliance · EU AI Act

The compliance posture covering CRA, RED, SBOM, and EU AI Act evidence. The audit manifest cross-links here.

See compliance →

SDK — developer surface

Extend the kiosk with custom MEP actions, MCP tools, and RAG corpus builders. Six-language SDK including Lua 5.4.

See the SDK →

All micro service components

Browse the full catalog of MOS4 micro services including mos-llm, mos-rag, mos-mcp, mos-voice, and mos-kiosk.

See components →

FAQ

Frequently asked questions

  • Does the kiosk require internet connectivity?

    No. The default posture is offline-first. The LLM, RAG, and voice all run on-device. Cloud access is opt-in and requires explicit configuration at three independent gates: system prompt, MCP tool allow-list, and device network policy.

  • Can the kiosk hallucinate?

    All language models can produce incorrect outputs. MOS4 kiosk addresses this structurally: the RAG refuse gate blocks answers when corpus similarity falls below threshold (top-1 cosine ≥ 0.55, top-3 mean ≥ 0.45). Below threshold, the kiosk answers "I do not have information on that" rather than generating a response. The per-answer audit manifest records what was retrieved, the similarity scores, and the refusal reason — so every answer is auditable.

  • Does it work with my existing hardware?

    MOS4 kiosk is silicon-agnostic above the 1 GB RAM floor. It targets compute-class and AI-class silicon. See the hardware page for the full silicon-tier list and form factors.

  • How does the EU AI Act posture apply to a kiosk deployment?

    The per-answer audit manifest emits provenance records suitable as evidence input for EU AI Act §10 (data governance) and §13 (transparency) obligations. See the compliance page for the full threat-model and posture documentation.

  • What is out of scope for MOS4 kiosk?

    Visual PPE/CV detection is covered by the AI Vision platform, not kiosk. Cloud orchestration and foundation-model training are also out of scope. The kiosk platform is purpose-built for grounded voice-first Q&A with an audit trail.

Bring your kiosk use case.

Noise environment, domain vocabulary, compliance requirements — bring the constraints and engineering will walk through the RAG, voice, and audit configuration.

Building on MOS4?

One reply from engineering, ~24h. No deck, no NDA.

Talk to engineering