DeOne

What it does

DeOne is a consumer dating product that scores compatibility from five psychometric assessments — Big Five, attachment style, communication patterns, connection languages, values alignment — and provides an AI relationship coach (Sage) who understands both users in the conversation, not just the one typing.

It lives at findd1.com.

We treat this as a safety-sensitive product, not a novelty chatbot. Compatibility scoring, relationship advice, and crisis language all need different guardrails.

How to inspect it

• Product leaders: Look at explainable matching and crisis handling before the chat polish. Trust and safety are product requirements here, not compliance afterthoughts.
• Engineering leaders: Inspect how scoring, retrieval, conversational context, and safety stops are separated. Each path needs a different eval because each failure harms the user differently.
• Operators: Ask what happens when a user expresses crisis language, when Sage lacks enough context, or when a match explanation is technically correct but emotionally wrong. Those cases shaped the release gate.

What we built, end to end

• The assessment engine and scoring model — five instruments, calibrated by the product owners and reviewed by their clinical advisor.
• The matching algorithm — vector similarity on 50+ derived traits, with explainable scoring so users can see why a match was suggested.
• Sage, the conversational coach — prompts grounded in the personality profiles of both users, not just one.
• The safety layer: crisis-language detection, escalation paths, hard stops on certain conversation patterns. Shipped before public launch — not after.
• The full consumer surface: onboarding flow, chat, profile visualizations, payments.

End-to-end build: 18 weeks across two phases.

What we wouldn’t ship until we got it right

• Crisis detection. A dating app that lets users talk about despair without a safety net is a liability the founder did not accept and neither did we. We delayed public launch by 4 weeks to land it.
• Explainable matching. Black-box compatibility scores erode trust in a domain where users are already skeptical. Every match shows the user why. This added two weeks; it is what we’d do again.
• The eval set for Sage. We hand-curated 200 conversation snippets across five tonal categories before we let Sage respond to a real user. The eval is in the repo. It is the only reason we shipped on time after the crisis-detection delay.

What we are honest about

• Early signal, small sample. The founder has shared two anecdotes publicly — Maya matched in 2 weeks, Jordan together at 6 months. That’s real data. It’s also not enough to claim a category-defining win, and the site says so.
• Dating is hard. Engineering is the part we can defend in any review.

Why this matters for your project

If your product depends on (1) scoring people or items across many fuzzy dimensions, (2) an LLM that has to read social context, and (3) the guarantee that a bad output cannot cause real harm — DeOne is the exact pattern. The eval discipline, the safety layer, the explainability work are all transferable; we have the receipts.