The problem
Teams increasingly draft with AI, but publish on their own reputation. Ask a single model whether something is true and you inherit that one model's blind spots — confidently delivered. For content you cannot afford to get wrong, one opinion is not verification.
The system
Fact Engineering treats verification as a courtroom, not a chatbot. Each claim in an uploaded document is cross-examined in parallel by a panel of frontier models from different providers — Anthropic, OpenAI, Google, DeepSeek, and others — and an LLM-as-judge weighs the evidence and issues one consensus verdict.
Scrutiny is dialed to the stakes across three tiers:
- Spell & grammar — editorial-grade cleanup in one fast pass.
- Model fact-check — the panel tests every claim against the sum of what it knows, then the judge adjudicates.
- Live research — every claim is re-investigated against the live web and returned with citations: defensible, traceable, audit-ready.
Engineering decisions that mattered
Asynchronous by design, real-time to watch. Fact-checking a long document fans out into many model calls. The pipeline runs on AWS Step Functions and Lambda, so checks scale without a fleet of servers, and results stream back to the browser over AppSync as each verdict lands.
One provider-agnostic model layer. Panel models are routed through a single gateway, so the panel composition — which models vote, who judges, what each tier costs — lives in one configuration file. When a better model ships, the product improves by editing one file, not by rebuilding the pipeline.
Boring where it should be boring. Authentication, payments, and storage use proven managed services — WorkOS, Stripe, S3, Aurora Serverless Postgres — so the engineering budget went into the part that differentiates: the consensus engine.
The outcome
A live, self-serve product: upload a document, watch a panel of frontier models argue about it, and export verdicts with sources. Credit pricing is quoted before every check, so the price you see is the price you pay.
For xternal AI, it is also the clearest possible credential. The advice we give clients — find the profitable bottleneck, build the smallest serious system, harden it into operations — is the same playbook we bet our own product on.