MEDIA PREVIEW
Overview & Result
Moritz Kremb built a real-time sales copilot application running as a desktop overlay alongside Zoom. By connecting real-time streaming audio from Deepgram to Jev, the system evaluates customer statements on the fly. When a customer raises pricing objections or mentions competitors, Jev classifies the objection type within 110ms and prompts the sales rep with the exact talking point before the customer finishes their sentence.
How Jev fits in the loop
- Captures live system audio and streams it to Deepgram Nova-2 for real-time speech-to-text.
- Every completed speech utterance is immediately sent to Jev to classify buyer stance (Budget, Competitor, Security, Authority, Closing).
- Jev returns the highest-scoring battle card ID and win-probability metric in 110ms.
- A sleek Electron desktop HUD displays the talking point, customer intent score, and suggested closing question in real time.
How to reproduce
- Follow instructions in Moritz Kremb's video demonstration.
- Launch local Electron app configured with virtual audio loopback to capture Zoom call speaker audio.
- Connect Deepgram Nova-2 streaming WebSocket for real-time transcript events.
- Observe Jev classifying objection types (Pricing, Timing, Competitor, Authority) and rendering battle-card prompts in under 150ms.
Why this build matters
Standard LLMs take 2-4 seconds to generate suggestions, which is far too late in a fast-paced live sales negotiation. Jev's sub-second classification makes genuinely ambient conversational assistance practical.
Reported performance
Reported by authorCost: $0.02 / 30-min call
Latency: 110ms / turn
110ms average latency from audio utterance completion to on-screen battle card prompt. Less than 2 cents total API cost per 30-minute sales meeting.
Limitations
- Requires clean system microphone audio without heavy room echo or dual-talk crosstalk.
- Sales battle cards must be curated and pre-loaded into the copilot registry for target products.