# CallCoach-AI

> Four agents score an investor call, and every flag has a ▶ Hear it button that plays the recording from the second it was said.

- Status: LIVE
- Result: WhipScribe Buildathon winner, Oct 2026, see 05 (https://whipscribe.com/buildathon)
- Period: Sep to Oct 2026
- Stack: Next.js, Flask, WhipScribe API, Groq GPT-OSS-120B, OpenAI / Anthropic / Ollama, MCP
- Live: https://callcoachai.sujit.top/
- Source: https://github.com/Blacksujit/callcoach-ai-whhipscribe
- HTML: https://www.sujit.top/missions/callcoach

## Objective

A founder raising a round takes 15 to 20 investor calls a week and almost never listens to one again. Feedback arrives as a feeling (*"that one went well"*), so the same weak answer survives a dozen pitches, and promises made on calls get forgotten.

## Baseline

No review at all, or a mentor who hears one call in twenty.

## System

- **Ingestion → WhipScribe** for transcription, speakers and timestamps.
- **Four scoring agents** (call quality, sales skills, objection handling, closing/compliance), each calling an LLM judge (Groq GPT-OSS-120B by default; OpenAI, Anthropic or local Ollama selectable). A **rule-based fallback** keeps scoring when no model is available.
- **Aggregator** joins scores, evidence and rules into one report; every flagged moment carries its timestamp.
- **Insight engine** for cross-call trends, commitments and action items, plus **Griot**, a chat that answers across all calls and cites the call and second each answer came from.
- **Delivery** to Slack, Notion and HubSpot. Each connection is tested with a real message before it's saved.
- **Same pipeline three ways:** web app, CLI (`python -m src.main --compare job1,job2,job3`), and an **MCP server** so Claude or Cursor can score calls directly.

## Eval

_Data pending: How often the four agents agree with a human mentor on the same calls; false-positive rate on compliance flags._

The design choice standing in for a metric: **evidence you can hear.** Every flag links to audio at the exact second, so the user verifies the model instead of trusting it.

## Anomalies

**A-1 · The host platform.** Building against WhipScribe meant reading its API and product closely, so I filed 67 findings on the buildathon repo: missing security headers, analytics firing before cookie consent, API responses drifting from the docs, pricing inconsistencies, accessibility gaps. [See the issues](https://github.com/neugence/whipscribe-buildathon/issues?q=author%3ABlacksujit)

**A-2 · Cold starts.** The analysis server sleeps when idle; the first request shows *"Waking up the analysis server…"* rather than hanging. Honest UX over a fake-fast spinner.

## Outcome

Live, no sign-up: open the app and click **Score a sample call**. Launch film and product walkthrough are linked from the repo.