Sous-Chef

The AI chatbot that converts visitors into briefed leads — live on this very site.

By the numbers

Results that moved the needle.

01 <0s
Time to first streamed token

02 -3
Inline rich components (services, plans, Calendly, lead badge)

03 $0
Avg. cost per full lead-capture conversation

Challenge

For a studio claiming “AI-Native”, a generic contact form isn’t a credential — it’s a liability. Visitors needed proof, not promises: a tangible AI experience that demoed the studio’s capability AND moved them through the funnel automatically. The team also wanted the chatbot to function as a portfolio piece, not just a service tool — every visitor interacting with Sous-Chef is also experiencing a live demo of what AppChefs builds.

Approach & Solution

We built Sous-Chef from scratch: a Claude Haiku 4.5–powered conversational AI agent embedded as a floating widget. It greets visitors, walks them through services with interactive inline cards (services grid, plan tiers, embedded Calendly modal), and uses Claude tool calling to capture qualified leads — emailing them automatically via Resend when intent is real. Sub-second streaming, a chef-themed personality (Cooking… Whisking… indicator), and a custom security layer (rate-limit, origin allowlist, input validation) keep cost and abuse in check. Prompt caching cuts repeat-input cost ~90%.

Under the hood

The actual pipes that power Sous-Chef.

Every message you send to Sous-Chef passes through this exact flow. Hover any node for the technical detail — the same architecture we deploy for client agents.

INPUT
User message
Visitor types a message in the floating Sous-Chef widget. Sent to /api/chat as JSON over HTTPS — response streamed back via Server-Sent Events.

GUARDS
Security gate
Per-IP rate limit (10/min, 60/hr), origin allowlist, hard input caps. Blocks abuse and API-key drain before the request ever reaches the model.

CONTEXT
System prompt + KB
Sous-Chef persona + AppChefs knowledge base (services, pricing, FAQs). Marked with cache_control: ephemeral — Anthropic prompt caching drops repeat-input cost ~90%.

LLM
Claude Haiku 4.5
Anthropic's fastest model. Decides at each turn: chat, call a tool, or emit a UI directive. Sub-second time-to-first-token thanks to caching + Haiku.

TOOL USE
capture_lead
When the model has name + email + brief in context, it calls this tool with a structured payload (service_interest, budget_range, timeline) — never invented.

DELIVERY
Resend
Sends an HTML email with the lead to maryl@theappchefs.com. Reply-to set to the lead's email so replies go straight to them. Also logs to leads.log.

INLINE UI
Component directives
Model emits {{component:services|plans|calendly}} tokens mid-stream. The widget parses them, replacing the directive with an interactive card the user can click.

STREAM
SSE → user
Server-Sent Events deliver text deltas, tool results, and lead_captured events token-by-token. The cooking-verb indicator (Whisking, Sautéing…) plays while waiting.

Tech Stack

Built with care.

Claude Haiku 4.5
Claude API
Tool Use
MCP
Vercel
AI SDK
Resend
React 19
Vite
SSE Streaming
Glass UI

Hey — I'm Sous-Chef, The App Chefs' assistant. I can walk you through our services, monthly Agency Plans, or help you book a discovery call. What are you working on?
See services
Build one like me
Agency Plans
Book a call