Rufus — Browser Runtime Demo

Edge → Cloud → Browser

The same Python step functions that run on a POS terminal or cloud worker run here, natively in your browser tab via Pyodide + WebAssembly.

📦

Order Fulfillment

STANDARD · PARALLEL fan-out · WorkflowJumpDirective

ValidateOrder → CheckInventory → FulfillmentDecision → ProcessPayment → SendConfirmation ⤷ BackorderNotice

Validates items and totals, fans out to two warehouse APIs in parallel, then either ships or places a backorder based on stock availability.

Result

📡

IoT Sensor Pipeline

STANDARD · loop logic · anomaly detection · health jump

InitPipeline → CollectSensorData → ProcessReadings → ComputeStatistics → HealthDecision ⤷ SendAlert

Generates 10 synthetic sensor readings with injected anomalies, computes statistics, and branches to an alert if the anomaly rate or σ is high.

Result

💳

Transaction Risk Scoring

STANDARD · WebGPU embedding · cosine similarity · decision jump

ExtractFeatures → GPUEmbedding → ComputeRiskScore → ScoreDecision ⤷ RecordOutcome

Model: all-MiniLM-L6-v2 (23 MB) not loaded

Downloading model weights (~23 MB) — cached after first run…

Embeds the transaction description using a sentence transformer model running on WebGPU (or WASM fallback), scores against risk pattern embeddings via cosine similarity, then approves, flags, or declines.

Result

📄

Document Summarisation

Qwen2.5-0.5B-Instruct · DECISION quality gate · extractive fallback

IngestDocument → PreprocessText → GenerateSummary → ExtractKeywords → QualityDecision ⤷ FallbackExtract

Model: Qwen2.5-0.5B (~350 MB, q4) not loaded

0 / 500 words

Model settings

Max tokens Rep. penalty Sampling Temperature 0.7

Downloading Qwen2.5-0.5B weights (~350 MB, q4) — cached after first run…

Abstractive summarisation using Qwen2.5-0.5B-Instruct (q4). A DECISION step validates compression ratio; if the output is too short it jumps to an extractive sentence fallback — a workflow-level quality gate.

Result

🔒

Air-Gapped Field Triage

bert-base-NER · On-device PII redaction · SAF routing · Zero-trust offline

CaptureReport → RunNERAnalysis → BuildRedactedPayload → RouteByPriority → LogStandard ⤷ EscalateIncident → StoreForForward

NER model: bert-base-NER (q8, ~27 MB) not loaded

Downloading bert-base-NER weights (~27 MB) — cached after first run…

A bert-base-NER model runs entirely on-device to strip person names and IDs from the raw report. The redacted payload is severity-routed by the DECISION step and written to local IndexedDB as a Store-and-Forward record — PII never touches the network.

Result

🧠

Paged Reasoning (wllama GGUF)

Q4_K_M GGUF via wllama · OPFS cache · Logic-gate fast path · real LLM output

AssessComplexity ⤷ FastPath | FullPagedInference → FormatOutput

Peak WASM RAM estimate —

0 MB300 MB limit (Safari)

Model:

88 MB Q4_K_M · works everywhere · fast

A complexity classifier routes simple queries to the fast path (first URL only). Complex prompts trigger full inference. SmolLM2-135M (88 MB) fits everywhere including Safari. Qwen2.5-0.5B (395 MB) gives higher quality on Chrome/Edge desktop.

Result

Run History 0

▼

No previous runs — run a workflow to start recording history

Worker Log