Incraft

I built Incraft after realizing parts of the pipeline behind Neurotype could... read more

• Pause-aware audio pipeline that inserts natural silence based on protocol timing, not just text breaks
• Scientific protocol-driven generation - ~50 evidence-based protocols (CBT, PMR, urge surfing) each with their own pacing rules, safety constraints, and structure
• Sessions shaped per protocol - a body scan breathes differently than a cognitive defusion exercise, and the pipeline respects that
• Every session is mixed and mastered - ambient layers, voice, and silence are balanced into a studio-quality output
• Intelligent sound selection that picks ambient soundscapes to match each session's tone and protocol
• Best-in-class TTS voices found by benchmarking 10+ providers to get the warmth, pacing, and naturalness right

Next.js for performance. The generation flow borrows ChatGPT's simplicity, the studio feels like Google Docs meets ElevenLabs. The script editor is built around original features that don't exist elsewhere: inline pause insertions, protocol block controls, and timing adjustments, all packaged in an ElevenLabs-style layout for maximum intuitiveness with zero learning curve.

Python/FastAPI on AWS Lambda (Docker via Mangum), Supabase, OpenRouter (Gemini 2.5 Flash). • 10-step script generation pipeline: validate, classify intent, select shape/sound, fill fields, write script, fix pauses, extend if too short, sanitize, and save. Steps that can run concurrently do via asyncio.gather • Born from optimizing Neurotype's pipeline. Pushed everything possible to precomputed JSON artifacts loaded at startup, making the pipeline deterministic where it can be and reducing LLM usage down to just 4 calls • Generate/render split: scripts are saved to Supabase after generation so users can edit them in the studio before sending to TTS • Separate TTS Lambda invoked via boto3 with a multi-provider setup (Inworld + ElevenLabs) behind a unified interface • Custom XML language created by the script generation pipeline is parsed by the TTS Lambda with pydub and ffmpeg for quick mixing and silences • Stripe for billing with webhook-driven subscription management. PostHog for analytics