Skip to content
← All work
AI product interface

Sightline

A real-time interview copilot: a dense control station for transcription, question detection, and answer shaping — with the ethics boundary stated on the page.

The Sightline session console: a dark two-column layout with resume and job-description pickers on the left, and session controls on the right including mode, interview type, difficulty, answer length, style, and a detection threshold slider set to 0.45.

2Tyme product

The problem

An AI product where the interface is the hard part. Speech arrives continuously, the model has to decide what is a question, and the operator needs to steer tone, length and framing without stopping to think.

The thinking

A control surface dense enough for real operators — segmented modes, a live detection-threshold slider, framework and domain switches — plus provider-agnostic model routing and a demo dataset so it can be shown without exposing anyone's data.

A session control panel: mode toggle between mock interview and live meeting, interview type, difficulty, answer length, style, a detection-threshold slider set to 0.45 with an explanation of how sure the system must be that something was a question, a standing-objective field, and framework and domain toggles.
The detection threshold is exposed, not hidden: 0.45 is how certain the system must be that it heard a question before it answers.

The system

Built as a TypeScript monorepo with the protocol and detector separated from the UI, and a provider abstraction over Anthropic, OpenAI and Gemini so no single vendor is load-bearing. Every control changes real behaviour — none of them are decorative.

Built with

 Realtime transcription
 Question detection
 Provider abstraction
 Session runtime
 Demo mode
StackTypeScript · React · Vite · Anthropic / OpenAI / Gemini

What it proves

A control surface dense enough for real operators — segmented modes, a live detection-threshold slider, framework and domain switches — plus provider-agnostic model routing and a demo dataset so it can be shown without exposing anyone's data..

What this is

One of our own products, running locally. The screen shows its built-in demo profile — fictional person, fictional company — not a real user's data.