Tools¶
Two interactive visualizations of sploink's architecture and bench results. Both are self-contained HTML — no JS frameworks, no external dependencies, white theme.
Architecture viewer¶
Open the architecture viewer →
Shows the workflow ↔ hardware bipartite assignment. Left column: the workflow IR (logical agent steps with data dependencies). Right column: available hardware types (CPU / LPU / GPU / Frontier), each containing the substrate instances (Ollama / Groq / Together / Anthropic) that serve them. Solid blue edges show the routing decision for the selected strategy.
Use it to:
- Understand the two-layer routing model visually
- Compare how
cpu_only/lpu_only/hw_routedstrategies redirect the workload - See which substrate instances are active vs planned
Regenerate locally: python -m sploink.architecture
Dashboard¶
Aggregates the latest bench results into a savings hero + bar charts + run history. Reads from bench/results/*.csv (CSVs produced by python -m bench.run).
The published dashboard above shows the n=30 HotpotQA run with the validated headline:
- 92.5% cost reduction (
hw_routedvslpu_only) - F1 trade-off of -0.132 with the current default policy (tunable; see Bench)
Regenerate locally with your own bench data: python -m sploink.dashboard
Switching between them¶
Both viewers include a pill in their header — click Architecture or Dashboard to switch instantly. The ← Docs link returns to this page.
How to embed these in your own docs¶
The HTML files are generated from a single python -m invocation each — no build pipeline required. Drop them anywhere your static site is served from:
python -m sploink.architecture --out public/architecture.html
python -m sploink.dashboard --out public/dashboard.html
They self-link assuming they're served from the same directory (the switcher pill uses relative URLs like dashboard.html).