Settings

Tune the actor-critic loop and the model it uses. Changes are validated, saved to the local database, and applied to the next run — no restart needed.

Loop knobs

Max refine iterations Extra refinement passes after the first draft (0–10). Score threshold Stop early once the weighted overall reaches this (0–100).

Target answer length (seconds) Spoken length to aim for, ~150 words/min (15–600).

Model

Generator model Choose your Bedrock model in the AWS Bedrock panel below — Load models from AWS, then Select model. The inference profiles (us./apac./au./global.) are the invocable ids for modern Claude/Nova. Critic model (optional) A sharper model for the critic only (same provider). Leave blank to reuse the generator model.

Provider

Active provider Which AI backend handles the next call. Switching takes effect on the next generate or test — no restart needed.

AWS Bedrock (no keys — ambient credentials) Bedrock authenticates with your ambient AWS credentials — no key is stored here. With both fields below blank it uses the default chain (env vars / AWS_PROFILE / SSO / instance role). If this box keeps credentials under a named profile with no [default], set the profile below. Then Load models from AWS and pick one — modern Claude/Nova models need a cross-region inference-profile id (e.g. us.anthropic.claude-…), which the list provides.

AWS profile (optional) Named profile from ~/.aws; blank = default chain. AWS region (optional) e.g. us-east-1; blank = from the environment / profile.

Loads against the profile & region typed above — no save needed. Save before Test connection / generating so those use the same config.

Voice

Have the interviewer read questions aloud (text-to-speech) and turn your spoken answers into text (speech-to-text). Choose the free in-browser voice or an AI provider, then Test it right here. Saved to the local database and applied to the next voice call — no restart. The Practice & Mock Interview pages use whatever you set here.

Text-to-speech — the interviewer's voice

Voice source In-browser is instant and free (your device's built-in voice). AI voice (ElevenLabs) is far more natural; the default eleven_flash_v2_5 model keeps latency low (~75 ms). If the delay ever bothers you, switch back to in-browser here. ElevenLabs API key (write-only) Stored locally and never shown back. Leave blank to keep the existing key; type a new value to replace it.

Lists the voices & models on your ElevenLabs account so you can pick from a list. Type a new key above first to list against it; otherwise it uses the saved key. No save needed to load.

Voice ElevenLabs voice id; the default is the "Rachel" voice. Model eleven_flash_v2_5 = low latency · eleven_multilingual_v2 = richer, slower.

Testing the AI voice saves these settings first, then synthesises a sample.

Speech-to-text — transcribe your answers

Transcription source In-browser is free but only works well in Chrome/Edge. AI model (Deepgram Nova-3) is the most accurate and works in every browser — your answer is recorded, then uploaded for transcription. Deepgram API key (write-only) Stored locally and never shown back. Leave blank to keep the existing key; type a new value to replace it. Model Deepgram model id; nova-3 is the current best-in-class default. Recording a test saves these settings first, then transcribes via Deepgram.