SaaS founder · @mertkilickaplan
Mert Kılıçkaplan
Founder background
Product Manager building WhichNotes | AI-powered knowledge graph for notes, links & documents | 0→1 SaaS & Growth
Expertise
- Product
Products
WhichNotes
Turn scattered sources into living, traceable research.
Recent posts
Ask WhichNotes now shows where the evidence ends.
One of the hardest parts of building an AI research product is deciding what should happen when the available sources don’t fully support an answer. Over the past two weeks, I moved Ask WhichNotes from “it returns cited answers” to a trust workflow that I’ve now verified in production. It can now: → distinguish source-backed, partially supported, and limited-evidence answers → show which individual claims are supported, weak, or unverified → explain why a cited source may not support a claim → open the original note behind the evidence → avoid inventing a conclusion when the required benchmark or data is missing Production testing also exposed two real defects: saved answers were carrying too many retrieved sources, and some structured answers timed out. I fixed both and re-tested the flow in English and Turkish, on desktop and mobile. This doesn’t prove that every AI answer will be correct. It gives WhichNotes a production-verified foundation for making uncertainty visible instead of hiding it. Next: improving language consistency, cleaning up follow-up questions, and observing how people use these trust signals in real research. When evidence is incomplete, would you rather receive a useful partial answer or a clear “not enough evidence” response?
The less visible part of shipping WhichNotes: making it…
This week I focused on the part of building an AI product that users rarely see: production hardening. WhichNotes turns notes, links, and PDFs into source-backed research briefs. That means trust is not only about answer quality—it also depends on data isolation, dependency safety, and predictable failure modes. Today I merged: → Reduced production dependency audit findings from 9 to 3, without forcing a risky downgrade → Added a real runtime smoke test for the language-detection pipeline → Made business-context isolation and database-function exposure audits required CI checks → Kept type checks, unit and integration tests, lint, production build, and preview deployment green This work is currently on the development branch and is being validated before the next production release. Next, I’m closing the remaining release gates and returning my focus to activation: helping new users reach their first useful, source-backed brief faster. For other SaaS founders: what “invisible” reliability work made the biggest difference before you pushed for growth?