Loading benchmark data…
EyesInAI·Loading live benchmark data
We scan your website and draft a complete assistant — bilingual, voice-capable, grounded in your own pages and live data. It answers on your site in one embedded line, runs behind hard spend caps, and measures its own conversations to route every question to the least-expensive model that still answers it well. You approve every change. Nothing goes live without you.
<script src="https://eyesinai.com/api/bot/your-business/widget.js" async></script>One script tag mounts a floating chat launcher on your site. Prefer your own UI? Call the API directly. Either way the assistant stays hosted and maintained here — config changes go live instantly.
Everything below is running in production today — nothing on this list is a roadmap item.
Give us your URL. We scan your site and draft a ready-to-review assistant — its name, tone, and a knowledge skill per topic (services, hours, location, booking), grounded in your own pages. You review and edit before anything goes live.
Every assistant answers in English and Spanish — it detects the visitor’s language and replies in it. More languages on request.
Turn on voice and visitors can speak to it and hear it back — speech recognition in, natural spoken replies out, with a typed fallback. Pick text-only, read-aloud, or full voice per assistant.
When a visitor asks about your catalog, prices, or availability, the assistant fetches live data from your systems (read-only) and renders it as a clean table — the numbers come straight from your data, never retyped by the model.
Follow-ups just work — "how much is it?" knows what "it" is. Slash commands give visitors one-tap shortcuts to the things they ask most.
Businesses get the same questions all day. A semantic answer cache serves repeats instantly at zero model cost — and refreshes itself automatically the moment your underlying data changes.
Point it at your own provider key and your account is billed directly (the key is held only as a server-side reference, never in your code). Or run on our metered, capped plan. Hard spend limits sit in front of every call, either way.
Drop one script tag on your site for a floating chat widget, or send visitors to a hosted chat page. Editing the assistant updates it live — no redeploy on your side, ever.
Visitors can hand it a document — a spec sheet, a request, a form — and ask questions about it in the same conversation.
The assistant measures its own real questions across candidate models and proposes better-value routing for your actual traffic — with the evidence attached. You approve; nothing re-tiers your live assistant without a human saying yes.
What actually happens between a visitor pressing enter and the answer appearing — and why every step makes the next answer better.
Every message passes safety gates before any model is involved: abuse and prompt-injection screening, rate limits, and hard spend ceilings. If a gate fails, no model is ever called — the caps are enforced before the spend, not after.
The assistant matches the question to one of its skills. Data skills fetch live, read-only answers from your systems; knowledge skills answer from your own pages. It answers about your business — it’s fenced from being a general-purpose chatbot on your dime.
The question’s task class picks the model: routine lookups go to a measured value model, hard turns go premium. If an equivalent question was answered recently, the cached answer returns instantly — no model call at all.
Live lookups render as exact tables; the model only summarizes around them. Voice replies are spoken. Every turn is logged with the model used, what it cost, and how fast it was.
A judge compares how well different models resolved your real questions — did it actually answer, with no invented facts. When a less expensive model provably holds quality, you get a proposal with the evidence. You approve it; the next conversation routes better.
Every hosted assistant is a tenant on Switchyard, our measured routing gateway — the same engine offered to teams that already run their own chatbot. Your assistant's questions are classified by task, and each task class carries its own measured model chain: routine lookups on a value model, hard turns on a premium one. The routing table isn't a guess — it's compiled from a benchmark that tests dozens of models daily, then tuned to yourtraffic by the nightly learn loop. The engine already runs a real company's customer and admin chatbots in production, where per-task routing cut model cost 31% with no measurable accuracy loss. Read the case study →
A public chatbot is an open door to your model bill. Ours ships locked down by default — these aren't add-ons, they're the floor.
Answers that appear as they’re written. Live stage updates ("checking availability…") already ship today.
Per-question economics: what your assistant answered, what it cost, and what the measured routing saved you against a single-premium-model baseline.
Your real questions (redacted, with your consent) become benchmark tests — so model rankings reflect your workload, not generic prompts.
Already run your own chatbot and just want the routing? Switchyard is the gateway on its own — change one line, keep everything else. Not sure either would help? Benchmark your use casefirst and we'll measure it honestly.
Give us your website and we'll come back with a working draft — name, tone, skills, and answers grounded in your pages. You see it before anyone else does.