02SaaS · AI Cold Outreach
Corpoo
Full Stack · .NET Backend · AI Pipeline · DevOps
B2B cold e-mail platform where an AI worker runs the conversation. Campaigns go out from virtual workers with their own persona, mailbox, quota and working hours; when a lead replies, the AI classifies intent, scores confidence and answers from the tenant's knowledge pack — above the threshold it sends, below it drafts or escalates, so a human is in the loop everywhere the AI is unsure.
.NETC#ReactTypeScriptLLMHubSpot API
Under the hood
FrontendReact · TypeScript
Backend.NET · C# · multi-tenant
AIIntent classifier · reply generation
GuardrailsConfidence thresholds · kill switch
CRMHubSpot · Webhooks · API
3AI decisions on every inbound reply
01
Lead replyinbound mail on a worker's thread
02
Guardrailskill switch · tenant pause · human takeover · suppression
03
Intent classifierLLM: intent, language, confidence
04
Decision engineSEND · DRAFT · ESCALATE by threshold
05
Knowledge packprohibited → approved facts → campaign → worker persona
06
Draft queuehuman edits, approves, escalates
07
Outcome + CRMthread closed, result synced to HubSpot
What's inside
- Virtual worker as the face of the outreach — name, persona, tone, signature, timezone, working hours, daily conversation cap and reply delay
- Lead search over an external company/people index with country, industry and seniority filters, plus CSV import and suppression lists
- Business campaign splits a segment across workers and their mailboxes, respecting daily new-mail quotas and the sending schedule
- Sending goes through an integrated e-mail provider — no separate account or inbox for the client
- Guardrails run before the AI: global kill switch, tenant pause, human-took-the-thread, closed thread, suppression — any one hands the conversation to a human
- LLM classifies intent and language of every inbound reply with a confidence score
- Decision engine picks SEND, DRAFT or ESCALATE by configurable thresholds; risk scoring can downgrade the decision
- Replies are written from the tenant knowledge pack in strict priority: prohibited claims → approved product facts and value proposition → campaign constraints → worker-scoped material → persona and tone
- Draft queue keeps everything the AI would not send: a human edits, approves, cancels or escalates
- Every decision is logged — intent, confidence, thresholds at decision time, risk flags, prompt and model version
- AI Overview: decision distribution, average confidence over time, drift, escalations and failures
- Inactive threads close automatically with an outcome (positive / neutral / negative) that feeds the worker's rating
- HubSpot sync for results, plus generic webhook and API paths for other CRMs
Screens
