
Most people never rehearse a salary negotiation out loud, and delivery is where they lose money: they pause before naming a number, correct themselves mid-sentence, rush when pushed back, or fold on the first counter. Negotiation Dojo lets you practice that moment for real. You speak with Alex, an AI hiring manager who makes an opening offer, pushes back, and only moves if you negotiate well. Under the hood, the Voice Agent API runs the live conversation and logs each of your moves (anchor, concession, counter-offer) through tool calls. The Voice Agent's user transcript has no word-level timing, so a second AssemblyAI Streaming STT v3 connection listens to the same microphone audio and returns per-word timestamps and confidence. Four detectors compare each answer against your own baseline: pause before a number, mid-sentence number retraction, pace spike, and mumbled number. When you end the call you get a Black Box style dossier: parameter rulers, a table of exceedances with your exact words, the single most expensive moment, and one concrete recommendation. The score is a fixed weighted formula (anchor 30%, reciprocity 30%, tell discipline 25%, outcome 15%). The LLM Gateway only writes the coaching prose and can never change the number. Deployed on Vercel and Render, MIT licensed, with 63 automated tests running in CI.
30 Sep 2026