Ingest
Firecrawl pages, Supadata transcripts, PDFs, and direct text enter one source model.
AI learning platform · Async pipelines / 2026
Explainer AI
Keeping slow, multi-provider generation reliable without blocking the product.
01 / The problem
Complexity should live in the system,not in the experience.
Audio and content generation cross several slow external services. Running that work inside HTTP requests would create timeouts and brittle user experiences.
02 / System flow
From input to a reliable outcome.
Generate
OpenRouter and Gemini create typed, structured learning outputs.
Orchestrate
BullMQ and Redis move podcast and conversational audio into background workers.
Persist
PostgreSQL tracks records, progress, credits, preferences, attempts, and ownership.
03 / Decisions that shaped the build
01Engineering choice
Checkpoint the pipeline
Scraped text, scripts, and audio URLs are saved after each stage so a retry can resume instead of repeating expensive work.
02Engineering choice
Treat preferences as prompt inputs
Reading level, tone, and depth are merged with each request rather than duplicated across output-specific UIs.
03Engineering choice
Structured state over opaque jobs
Users see persisted progress and actionable failure states while workers run independently.
04 / Reliability
Designed for imperfect conditions.
- Two-attempt queue retries with backoff
- Ownership checks and JWT-protected resources
- Rate limiting, Helmet, and SSRF filtering
- Transactional credits and refunds on persistence failure
05 / Reflection
“The project shifted from a set of AI buttons into a system: shared ingestion, reusable source context, durable jobs, and output-specific experiences.”Explore selected work