coastlight

Practice Areas

We work with product, research, and engineering teams on the parts of AI development that don't fit cleanly into a benchmark. Our focus is the methodology and judgment that turn responsible AI from a stated principle into shipped behavior.

• Evaluation methodology
  Designing evaluation sets, metrics, and annotation processes for generative and predictive systems. Subjective measures operationalized into rigorous comparison tasks. Demographically balanced test sets, inter-annotator calibration, and reproducible reporting.
• Red-teaming and harm analysis
  Adversarial probing for generative and multimodal systems, including learned associations, edge-case behavior, and risks specific to high-stakes deployment contexts. Practical mitigations that go beyond keyword blocklists.
• Governance and review
  Cross-functional review processes between ML, product, legal, and ethics teams. Data provenance documentation, datasheets, and pre-launch frameworks that hold up to internal and external review.
• Responsible deployment
  Pre-launch readiness, post-launch monitoring, and messaging strategy for systems with predictable failure modes. We work alongside trust and safety, legal, and product teams rather than handing off a checklist.

Experimental Applications

Alongside the consulting practice, Coastlight builds small applications that explore how thoughtful AI workflows can benefit individuals and the people around them. The apps are independent products and a working sandbox for the methodology questions we work on with clients.

• AI Writing Guardrails↗
  A process enforcement skill for AI-assisted academic writing. Surfaces literature bias, gates the workflow at decision points, and keeps an honest record of how the AI was used.
• Virgil In development
  A voice agent for personal context, memory, and the conversations that help you think out loud about what matters. Currently in active development.
• Harbor Habits↗
  A daily habit tracker organized around the life areas you actually want to maintain. Built MCP-native so it integrates with the assistants people already use.
• Surge Flow↗
  Focus support for the day's real work. Breathe, clarify, flow.

Who we are

Coastlight is led by Jay Weiler. His work over the last decade has focused on evaluation methodology, governance, and the cross-functional review that turns responsible AI from a stated principle into shipped behavior. He works on generative and predictive systems, and on the product questions that surround them.