Project Acme — Architecture

Recreated outline. Original was lost to the SAMPLE_WELCOME race.

Overview

Acme is a post-call sales intelligence layer. Components (high level):

Copy
audio in → transcription → extraction → CRM write
            (Whisper)       (Claude)     (Salesforce / HubSpot)

Components

Ingestion

• Cloud upload endpoint (POST /v1/calls)
• Webhook from Zoom / Gong / Chorus
• File-store: object storage (raw audio, retained 30d)

Transcription

• Provider: Whisper (self-hosted) or Deepgram (managed)
• Output: speaker-diarised transcript JSON

Extraction

• LLM: Claude with structured output (JSON schema)
• Inputs: transcript, account context, recent CRM state
• Outputs: follow_ups[], decisions[], next_steps[]

CRM bridge

• OAuth tokens per tenant (Salesforce, HubSpot)
• Idempotent write with external ID = call_id
• Approval queue: nothing writes without human OK in v1

Data model

• calls(id, tenant_id, audio_url, status, transcript_id, ...)
• extractions(id, call_id, kind, payload, status, approver_id)
• integrations(tenant_id, provider, refresh_token, ...)

Open questions

• Streaming vs batch transcription
• Multi-tenant inference vs per-tenant model pinning
• Retention: raw audio vs transcript vs extracted-only