✦ SELLERIUM — TAILORED DEMO ✦

GTM Systems AI Agent — live demo

Built for the Senior GTM Systems AI and Automation Engineer role (Camunda, fully remote): an AI agent that answers go-to-market questions from CRM data — with retrieval, guardrails, a tool loop, and an audit trail, exactly as the role describes.

Honesty note — please read first: every record below is fictional sample data in a Salesforce-shaped schema (accounts, opportunities, contacts). No real Salesforce org is connected. The Salesforce connector is a planned ramp, not past production work — the team has deep API-integration experience with other business systems, and this demo shows the agent layer (retrieval, guardrails, tool loop) that transfers across CRMs. Retrieval and guardrails on this page run for real in your browser; the final answer wording is template-composed from the retrieved records (no live LLM key — labelled wherever it appears).

1) Ask a GTM question

Try a suggested question, or type your own. Try the last chip to see the guardrails block an attack.

Question

3) Sample CRM dataset fictional data — Salesforce-shaped schema

Objects mirror Salesforce: Account (Id, Name, Type, Industry, AnnualRevenue, Owner, Health_Score__c), Opportunity (Id, Name, AccountId, StageName, Amount, CloseDate, Probability), Contact (Id, FirstName, LastName, Email, AccountId, Title). Domains use .example so no real person is involved.

Accounts

Opportunities

Contacts

4) Tools (MCP-style) & guardrail rules

Tools the agent can call

Tool	What it does
`crm.lookup_account`	Find an account by name
`crm.list_opportunities`	Open opportunities for an account, optionally above an amount
`crm.churn_scan`	Flag accounts with health < 50 or a stalled (past close date) open opportunity
`crm.draft_next_action`	Draft the next step from the opportunity stage playbook

Schemas mirror the MCP-server pattern from the role posting. Here the tools run against the sample dataset in your browser — in production they would call the CRM API.

Guardrails (run before any model sees text)

PII redaction: emails → [EMAIL], phone numbers → [PHONE], card numbers → [CARD]
Injection screen: blocks “ignore previous instructions”, “disregard instructions”, “system prompt”, “you are now…”, “pretend…”, “jailbreak” and similar
Confidence threshold: top retrieval score below 0.10 → answer is withheld, routed to a human
Human escalation: complaints, and high churn risk, route to a human owner (demonstrated, never sent)
Audit logging: every run is logged below with timestamp, tools called, and outcome
Cost / latency: retrieval runs client-side — $0.00 LLM cost; latency is measured live per run

5) Audit log real — appended on every run

No runs yet. Run the agent above and each run will be logged here.

6) What this proves — and what it does not

✓ Proven here

Retrieval pipeline (TF-IDF + cosine) over structured CRM records
Guardrails: PII redaction + prompt-injection screen, executed for real
Tool loop: lookup → opportunities → next action, with sources cited
Workflow rules: churn scan, stage playbook, escalation, audit trail
Cost/latency transparency: measured latency, $0 LLM cost client-side

✗ Not claimed

Production Salesforce work — no real org is connected (planned ramp)
GCP deployment — this demo runs fully in the browser
Production LLM evaluation at scale — generation here is template-composed
Any customer metric — all figures are fictional sample data

Transferable production experience behind this demo: 24/7 AI secretaries on WhatsApp/Telegram/Instagram (maryam24.ir), dispatch automation (etesal24.ir), n8n workflow automation, and LangGraph tool-loop agents — see the Sellerium demo gallery.