All work

Sales call to CRM sync

Turn transcripts into grounded CRM diffs and follow-up drafts, with approval required for every write.

TypeScriptNode.jsPostgreSQLDocker ComposeGitHub Actions
Recorded CI output showing a proposed CRM diff, review flags, approval and an already-applied retry.
Recorded CI output. Open the source log ↗

This is a command line implementation with recorded runs. How to read this evidence.

Tests and demos use a deterministic mock provider. A Claude adapter sits behind the same interface; its live quality remains unmeasured.

Problem

Calls contain multiple speakers, missing facts and contradictory budgets or dates. A CRM update needs traceable evidence and review of the exact change to an existing record.

Architecture

  1. 1. Extract with provenance

    A mock or Claude provider proposes supported fields with verbatim quotes and speaker labels. Field-specific checks reject unsupported evidence.

  2. 2. Merge all chunks

    Long transcripts are processed in chunks. Overlap is merged by source statement, while separate conflicting statements remain flagged.

  3. 3. Show the exact diff

    PostgreSQL stores the before/after proposal, evidence, ambiguity flags and base CRM version. Missing fields preserve existing values.

  4. 4. Approve and apply

    The trusted operator approves a canonical diff hash. An optimistic version check, CRM write and audit commit atomically; repeated calls or applies reuse checkpoints.

Tradeoffs

  • Grounding deliberately uses a limited English/Urdu phrase set. Unsupported language or values require review.
  • Next week means the next Monday in the call timezone and is flagged as an interpretation.

Test results

30 distinct test cases passed in at least one appropriate environment. Zero failures in the linked CI run.

Actual results from the linked GitHub Actions run
EnvironmentPassedSkipped
Native PostgreSQL300
Embedded PostgreSQL300

No test cases were skipped in either environment.

Tests cover separate speakers, shared-company contacts, distant conflicts in a 4,000-line transcript, stale approvals, lost work and repeated applies.

Docker Compose built and ran the seeded demo from a fresh Actions checkout.

Verified code: cc1bc499c73958e810218b9baee00d2c6fd3f8e0

Recorded demo output

Docker stdout from the linked CI run.

All data synthetic. No external CRM writes or messages.
Review diff: {
  "contacts": {
    "after": [
      "Amina",
      "Bilal"
    ],
    "before": []
  },
  "nextStep": {
    "after": "send proposal next week",
    "before": null
  },
  "budgetMinor": {
    "after": 1000000,
    "before": 500000
  },
  "nextStepDate": {
    "after": "2026-01-26",
    "before": null
  }
}
Flags: [ 'crm-contradiction:budgetMinor', 'vague-next-week:2026-01-26' ]
Follow-up draft: Synthetic follow-up draft. Proposed next step: send proposal next week. Review flags: crm-contradiction:budgetMinor, vague-next-week:2026-01-26.
Before approval CRM version: 1
Dry run: dry-run
Explicitly approved apply: applied
Retry: already-applied
Audit records: 1
Read the recorded output in the repository

Known limits

  • The CRM is a PostgreSQL-backed mock, with no real CRM writes or follow-up messages.
  • Speaker labels must already be present; there is no audio transcription or diarization.
  • Transcripts are limited to one million UTF-16 code units. Live Claude quality is unmeasured.
Read policy and implementation assumptions
Back to portfolio