TheBeat

The specialist tech-news space is noisy and slow, and most “AI news” is either shallow aggregation or unverifiable summarising. The aim was a greenfield platform that could ingest a high volume of sources, enrich and organise them, and surface genuinely useful, grounded insight, without hallucinating or drowning readers in duplicates.

A resumable ingestion and enrichment pipeline feeds a retrieval-augmented experience. Content flows through a staged, checkpointed pipeline so long runs resume rather than restart; a cost-aware model router sends each job to the right model; and every AI-surfaced claim is grounded in retrieved source material. Evaluation and guardrails are part of delivery, not bolted on afterwards.

Keeping generative output trustworthy at volume, controlling inference cost without capping quality, and building the whole thing AI-first (spec-driven, agent-executed) while holding a firm quality bar through review and evals.

A running, greenfield platform built and operated primarily through agentic workflows, with the measurement layer, golden sets, regression gates and tracing, that makes its output safe to trust. The underlying patterns were open-sourced as agent-eval-starter and resumable-llm-pipeline.