← back
saas · llm

Bylgja

2025 In private development

What it is

Bylgja is a multi-product SaaS platform built for digital publishers and subscription businesses. It ships four products on a shared technical core: Chat (LLM chatbots), Search (hybrid semantic search), Curate (content recommendations), and Pulse (newsletter intelligence). Each product is deployed via a single script tag with no build step required on the customer side.

Problem

Media companies and subscription businesses in B2C face several high-value problems that are individually hard to solve: engaging readers with relevant content, surfacing it through intelligent search, retaining subscribers through personalized recommendations, and reducing churn before it happens. Each problem has its own tooling ecosystem, its own vendor, and its own integration cost.

Bylgja addresses all four from a unified platform - shared authentication, shared billing, shared infrastructure - so a publisher can adopt one product and expand to others without re-onboarding. The alternative is stitching together four separate SaaS vendors, each with its own data model, API, and contract.

What’s innovative

All four products deploy through a Shadow DOM-isolated script tag - one line of HTML, no framework dependency, no conflict with the host site’s CSS. The shared backend means customer data flows across products: content ingested for Search is also available to Curate; subscriber engagement signals from Pulse can inform recommendation surfaces.

Behavioral and interest representations are built without cookies or PII, making the platform compliant by design under GDPR and ePrivacy. The LLM layer in Chat is configurable per customer: local Ollama for privacy-sensitive deployments, OpenRouter for cloud access.

What I built

Shared core: FastAPI backend with async PostgreSQL, Redis for caching and rate limiting, Celery for async workloads, and a Next.js 15 admin portal covering all four products.

Product layers on top:

  • Chat: per-org Qdrant collections, WebSocket token streaming, RAG pipeline, conversation storage
  • Search: hybrid BM25 + vector retrieval with configurable weighting, re-ranking, query analytics
  • Curate: per-reader interest vectors from reading behavior, TimescaleDB for event storage, hot recommendation caching
  • Pulse: behavioral churn scoring from engagement signals, TimescaleDB time-series storage, campaign orchestration (in progress)

This project is in private development. Want a walkthrough? Reach out on LinkedIn.