Guardrailed Outreach Agent

LLM Workers Behind Deterministic Rails

Demo mode: the prospect is fictional and never crawled. Cadence is a made-up product. The pipeline drafts outreach but never sends — it stops at human review. Deterministic guardrails wrap every LLM step.

Atlas Pottery Studio

Independent ceramics studio & class provider · Mara (Owner) · ★ 4.8 (213)

Atlas Pottery Studio runs beginner and advanced wheel-throwing classes, open studio sessions and private events. To book a class, call the studio during opening hours — our schedule fills up fast, especially for weekend slots. Run by Mara, a working potter, since 2014.

signal 1signal 2signal 3signal 4signal 5

Tamper test: drop the opt-out line (demonstrates the critic's hard-fail short-circuit)

Guardrailed Outreach Agent

Case Study

Industry

B2B Sales / Outbound — personalised cold outreach at scale

Context

Built for a product that wants to reach small businesses with genuinely personalised outreach — not spray-and-pray. The risk is obvious: an LLM let loose on scraped data will happily write something inaccurate, creepy, or non-compliant. This demo uses a fictional prospect and a made-up sender, and never sends anything.

The Challenge

A free-roaming “AI SDR” is a liability. You want the personalisation an LLM gives you, but with hard guarantees: no private data in emails, no over-claims, always an opt-out, always a human approval before send.

Pain Points

✖LLMs personalise from whatever they're given — including private or creepy details

✖Cold email has legal requirements (opt-out, sender identity) an LLM can silently omit

✖“Agentic” systems are hard to audit when something goes wrong