Building a tool that turns client calls into ready-to-send proposals

Описание и применение

Инструмент Bespken автоматически присоединяется к клиентским звонкам, транскрибирует их, извлекает ключевые данные и генерирует готовый к отправке коммерческий предложение. Однако низкая уверенность в числовых данных требует обязательного ручного подтверждения.

Задача

Freelancers and consultants often need to reconstruct scope, pricing, and timelines from memory after a client call, leading to errors and inefficiencies.

Результат

Bespken is fully functional end-to-end, producing proposals from call transcripts. However, transcription confidence scores are low, especially for numbers, necessitating a mandatory review step.

Рабочий процесс

• Connect Google Calendar via Supabase Auth + Google OAuth (calendar.readonly scope).
• View upcoming calls with video links on a dashboard.
• Click "Join with Bespken" to add a bot participant via the Meeting API.
• When the call ends, a webhook triggers and fetches the transcript from a presigned URL.
• Send the transcript to Gemini for structured extraction of client, scope, price, timeline, and notes.
• Display extracted fields in a review screen with editable inputs and a prominent warning to verify everything.
• Generate a proposal document from the (edited) data and present it for sending.

Доказательства

This is fully working end-to-end in production right now, not just a mockup.

Источник

Marathi test call: failed badly. Confidence scores mostly under 0.5, numbers specifically got mangled.

Источник

English test call: numbers came through textually correct, but confidence scores were still surprisingly low (0.01–0.4) even on correct text.

Источник