Jev call copilot for CurrencyTransfer
Working documents for a live call copilot built on Cloudflare Workers and TypeSafe's Jev model, modelled on the open-source sales copilot and adapted for onboarding and customer-success calls. The deliverables come first: the demo script and the measured reports. Then the architecture; the research documents are its evidence.
Documents
Deliverables
- Demo script
The five-minute demo: one onboarding and one customer-success call, every step cited by utterance number and player clock.
- Latest evaluation run
The newest committed evaluation run: replay checks against the labels, cost and latency gates, per-call results.
- Measured cost and latency
Every Jev request of a dedicated cost run: tokens, latency, cost per replayed call, and the AI Gateway cross-check.
- Speech-to-text report
Deepgram nova-3 on Workers AI: neurons and transcription time measured per recording.
Plan
- Milestone M1b PRD for your review
Forty one-commit stories for live granularity, card lifetime, the adaptive plan and pre-call facts, checked adversarially five times. Review before it is converted and run.
Review
- Your replay review of 27 September and the answers
Turn-level updates, the plan that read like a script, the vanishing suggestion, the double tick, the false rate concern and the jumping scrubber: what the data shows and what changes, with the decisions that are yours.
- Proposal M1b: live granularity and the adaptive plan
The engineering proposal behind the 27 September answers: sentence-level decisions in replay, card lifetime rules, plan items as facts, must-says and judgement calls, pre-call facts, and 22 one-commit stories.
- Rate objections: your playbook in the system
How the five objections, the qualifying questions and the approved lines are encoded as situations, and what the copilot shows for each.
- Like-for-like: situations and client examples
Every approved line sits under a situation with real client turns from the harvest; a point only shows when the client's turn fits.
- Labels, headlines, hints and caution notes
Every label the dashboard shows around the approved lines, all approved on 26 September; reword any by id and it ships as the next policy version.
- Red-pen sheet: playbook lines and situations
Every line and situation point the copilot may show a rep, generated from the policy JSON with its approval status.
- Red-pen sheet: real rep lines per move
For each move, the phrasings real reps used on the call-coach corpus, clustered, with call counts and acceptance rates.
- Dashboard design system
The written spec of the approved design G: tokens, layout, panels, the living call plan and the client facts chips.
- Your design review and the answers
Your 25 September review of the first variants and what changed because of it.
- Round 3 design brief
The concept of the living call plan, what the screen must do, and the anti-patterns from your review.
Architecture
- Architecture
The synthesised design: components on Cloudflare, data model, the per-utterance decision loop, curation, evaluation, compliance, and milestone 1.
- Proposal A: fastest proof of concept
Least machinery to get milestone 1 in front of you.
- Proposal B: production-shaped
The shape the system needs in production, with milestone 1 as its first vertical slice.
- Proposal C: data-first
Curation and evaluation loop first; the live dashboard consumes corpus-derived policy.
Research
- Reference copilot dissection
How the open-source Jev sales copilot works, line by line, and what transfers.
- CurrencyTransfer domain brief
Customer journey, product facts, draft question banks and playbooks for onboarding and customer success. Marked DRAFT for review.
- Corpus and curation
Using the historical call corpus: ingestion, Jev quality rubric, the admin curation loop, storage, evaluation harness.
- Speech-to-text options
Whisper vs Deepgram nova-3 on Workers AI, diarisation, costs, and the live verification on a real recording.
- Prior art
Other Jev call copilots, commercial coaching tools, and the ten most transferable ideas.
Reference
- Jev builder's guide
Request and response contract, semantics, Cloudflare integration, patterns, limits and gotchas. Verified against the vendor docs.
Verified so far
Decisions taken
- Milestone 1: upload or pick a recorded call and replay it against a live dashboard: signals, stage, must-say checklist, objection handling, next best move.
- Both scenarios from the start: onboarding (the activation call after the sign-up wizard) and customer success.
- Headline metric: must-say completeness plus risk flags for onboarding; resolution and call health for customer success. No closing probability.
- Suggested lines: pre-approved playbook lines, plus tailored rewrites when Jev verifies them for invented facts, promises and rate predictions.
- Curation: Jev scores every historical call; the admin marks Jev-tagged moments as model, acceptable or avoid.