The AI industry’s obsession with larger parameter counts has hit a wall of diminishing returns for enterprise reliability. Skeptical, a new startup founded by researchers from KTH Stockholm, has raised pre-seed funding to solve a different problem: not making the AI smarter, but making its mistakes predictable. The company’s core product routes only the 'hard cases' to human review while automatically clearing the rest, all within a statistically guaranteed error budget that users define themselves.

The 'No Retraining' Promise

Unlike most AI middleware that requires fine-tuning or model swapping, Skeptical operates as a post-hoc verification layer. The system ingests the predicted scores from your existing model and compares them against correct answers from past reviews. It then calculates a confidence threshold that ensures the error rate among cleared decisions stays at or below your specified budget with 95% confidence. This approach means a weaker model simply clears less volume, but the reliability of what does get automated remains mathematically bounded.

Peer-Reviewed Roots and Leadership

The technology is grounded in research published at VLDB 2025 and VLDB 2026, lending it the academic rigor often missing from 'vibe-coded' AI tools. The founding team includes CEO Sonia Horchidan, a former Googler with a PhD from KTH on trustworthy AI-native data systems, and Chief Science Officer Paris Carbone, an Associate Professor at KTH who pioneered modern data streaming and received the ACM SIGMOD Systems Award for Apache Flink. Fabian Zeiher, an MSc graduate in distributed systems, serves as the founding engineer, building the production stack that keeps these bounds stable even as data distributions shift.

Auditable Automation for the Enterprise

For compliance-heavy industries, the ability to put a number in front of an auditor is the killer feature. Skeptical provides a statistical bound that holds true over time, using a small audit sample of cleared decisions to feed recalibration loops. This allows companies to automate high-volume tasks without the 'black box' anxiety that often stalls AI adoption in regulated sectors. The company offers a free 'Automation Score' browser tool to let teams test their data against these bounds before committing to a full integration.

Key Takeaways

  • Skeptical raises pre-seed funding to enforce statistical error bounds on AI outputs without model retraining.
  • The system guarantees error rates stay within user-defined budgets with 95% confidence, based on peer-reviewed VLDB research.
  • Leadership includes KTH researchers and ex-Google engineers, emphasizing academic rigor over hype.
  • Integration is non-invasive, reading only predicted scores and past correct answers from existing pipelines.

The Bottom Line

This is the anti-hype product we’ve been waiting for: reliability engineered through statistics rather than brute-force scaling.