Discovery Call Analyser v2

An AI coach for SDRs and outbound sales teams. Role-play a discovery call or chat with an AI buyer, or add a conversation of your own, and it:

• Scores how well the lead was qualified with MEDDICC, BANT or SPICED, quoting the conversation as evidence for every score
• Flags what wasn't asked and gives the exact questions to ask on the next call
• Shows what to say better: quotes the rep's weak lines and the buyer cues they missed, gives the exact words to use instead, and names the one thing to fix first
• Recommends a verdict (Advance, Nurture or Disqualify) and next steps
• Highlights the transcript, tagging each quote with the element it supports
• Drafts the follow-up email to the buyer, ready to copy or open in your email app
• Writes CRM notes, ready to copy or export with your history as CSV
• Saves a PDF report of the scorecard, transcript, email and CRM notes
• Tracks progress: average score over time, the elements you miss most often, and the selling skills that come up most in coaching

Conversations can come from a role-play with an AI buyer, or be pasted, uploaded as a .txt file, uploaded as an audio recording, or recorded in the browser. Audio is transcribed on the user's own device, so recordings are never uploaded.

Version 1, built for my Business Analyst with Applied AI course, is at discovery-call-analyser. This is a clean rebuild with audio input, three frameworks and a new design.

What changed from v1

Area	v1	v2
Input	Paste or .txt	Paste, .txt, audio file, or record in the browser
Frameworks	MEDDICC	MEDDICC, BANT, SPICED
AI output	JSON parsed from free text	Forced tool call with a JSON schema, then validated
Demo mode	Saved MEDDICC results for 5 calls	Saved results for 6 conversations in all 3 frameworks
Practice	None	Role-play calls and chats with four AI buyers, live hints, and a reveal of what the buyer knew
Outputs	Scorecard, CRM notes	Adds "Say it better" line-by-line coaching, highlighted transcript, follow-up email draft and PDF report
Progress	History table	History plus a dashboard: score trend and most-missed elements
Tests	None	Automated tests for validation, scoring bands and the API

How it works

flowchart LR
  A[Paste, .txt, audio file<br/>or browser recording] -->|audio only| W[Whisper speech-to-text<br/>runs in the browser]
  W --> B[Editable transcript]
  A --> B
  B -->|POST /api/analyze| C[Serverless function<br/>Vercel]
  C -->|framework rubric + JSON schema| D[Claude<br/>Anthropic API]
  D -->|record_analysis tool call| C
  C -->|validated scorecard| E[Web page: verdict, scores,<br/>questions, CRM notes, CSV]

1. Audio is decoded in the browser, split at quiet moments into one-minute parts, and transcribed by Whisper (whisper-base.en) running on the device with Transformers.js. The first run downloads about 80 MB of model files, which the browser then keeps. Whisper doesn't identify speakers, so the prompt tells Claude to work out who is the rep and who is the buyer, and the user can add names before analysing.
2. The server adds the chosen framework's rubric and forces Claude to answer through a record_analysis tool whose schema fixes the number of elements, the 0–2 score range and the three verdicts.
3. Validation (validate() in api/analyze.js) checks every field again before anything reaches the page, so a malformed response can't break it.

Role-play with an AI buyer

Reps practise on the page instead of on real prospects:

1. Pick a buyer, a channel and a difficulty. Four fictional buyers, each with a different challenge: an open but busy finance head, a guarded CIO, a price-sceptical founder, and a tough VP who is already evaluating two vendors. Choose Call (talk into the microphone or type, and the buyer can answer out loud) or Chat (typed, WhatsApp style), and Friendly, Realistic or Tough. Reps can enter their own name and what they sell.
2. Run the conversation. Each buyer has private facts (pain, numbers, budget, who signs off, decision process, paperwork, timeline, competition, a possible champion and objections). They only share a fact when the rep asks a question that earns it, push back on early pitching, and only agree to a next meeting if the rep has earned it.
3. Press Hint when stuck. A separate coach reads the conversation so far (without seeing the buyer's private facts) and suggests the exact next line, tagged with the framework element it targets.
4. End and score. The conversation goes straight into the analyser and opens on Say it better. The page then shows What the buyer knew, so the rep can see which facts they never uncovered.

How it's built: api/roleplay.js keeps the buyer personas on the server (public/personas.js), so the page can't change how a buyer behaves, and the rep's "what you sell" text is passed to the model as data, not instructions. Replies use a fast model so the conversation feels live (see AI models below). In call mode the rep's speech is turned into text on the device with the same Whisper model used for uploads, and the buyer's voice uses the browser's built-in speech synthesis.

Scoring

Each element is scored 0 (not covered), 1 (partly covered) or 2 (clearly established).

Framework	Elements	Max	Best for
MEDDICC	Metrics, Economic buyer, Decision criteria, Decision process, Paper process, Identify pain, Champion, Competition	16	Enterprise deals with several approvers
BANT	Budget, Authority, Need, Timeline	8	Quick first-call checks
SPICED	Situation, Pain, Impact, Critical event, Decision	10	Value-led SaaS discovery

This is the beginning of the README. Read the rest at the original. Read the full README ↗