A member posts a need, a curator proposes an introduction with a written reason, and both people approve before any contact detail moves. The whole loop works end to end. It is the software part of a bigger idea I am building with a friend, and one of the ways I keep getting better at technical product management without waiting for anyone else.
By Nova Skye, directing AI-assisted development.July 2026 – present5 minute read
Stage
Working end to end · the software part of a bigger idea
My role
The thesis, the starting focus, the north-star metric, the privacy model, the process, every scope decision, and the approval of plans before implementation
Professional networks show where someone has worked. They do not show what that person needs right now, what they can credibly provide, whom they trust, or whether they are available. Business owners in a financial-media audience kept asking the same question, "do you know a good fractional CFO?", and the honest answer depended on one person's memory.
I started with one group, finance and operations professionals for business owners, so one full introduction loop could be tested end to end.
2
The bet
The hypothesis is that a trusted intermediary who proposes introductions with a written reason, and who records what came of them, will produce more verified valuable outcomes than an open directory.
The success signal is verified valuable outcomes per 100 active members, not sign-ups or profile counts.
3
What was built
Members state availability and post a structured need: engagement type, timeline, budget band, location, remote or not. The curator sees candidates ranked by a deterministic matcher with plain-language reasons, proposes an introduction, and both parties approve before contact details move. After acceptance each side can privately rate the introduction, a reminder goes out after two weeks, and a "needs attention" page surfaces anything stale.
A curator proposing an introduction. The reason is required. "~fundraising" marks a related-term guess so it is never mistaken for an exact match. Tap to enlarge.Fig. 01
Inviting a standing cohort, ranked by the same engine as needs, with no new matching code. Tap to enlarge.Fig. 02
Consent is captured per purpose and unbundled, at first sign-in. Deleting an account erases the person's data and withdraws their live introductions, while the record of their consent and of the deletion itself stays. Before any events schema was written I had the formats of comparable products researched: each one ran a standing cohort, the same eight to sixteen people on a cadence.
4
The decision log
No.WhenDecisionWhy, and what was rejectedWhat followed
R-01Jul 2026Start with one narrow service areaSo one full introduction loop could be tested end to end. Rejected: a general profile database, because more profiles would not show whether introductions work.The full introduction loop was built end to end.
R-02Oct 2026A curated synonym table for ranking, not an AI callRanking runs on every page load for every candidate. A model there adds cost, latency and a failure mode, and stops being unit-testable. The table costs upkeep instead.The same ranking engine was reused for curated events with no new matching code.
R-03Oct 2026Deletion erases the data and keeps a record that it happenedThe person's data has to go, and the consent and audit history has to stay.Deleting an account now clears personal data and withdraws live introductions, and the consent and audit records stay.
R-04Oct 2026One brief per feature, a PR checklist, a CI gate, a decision register, and a plan before any major upgradeDocumentation has to keep up with the code, so the check runs at the moment of change. A major upgrade needs a plan with risks, tests and a way back, tried first on a copy.The gate fails any PR that touches application code without a brief; it failed a real PR as designed, then passed.
5
The result
Outputs: what shipped
The full introduction loop: needs, ranking with reasons, mutual approval, outcomes, private feedback, reminders, an attention queue
Consent per purpose, and account deletion with an audit record
Curated events as standing cohorts, on the same ranking engine
A CI documentation gate and a deploy build in CI
Outcomes: what changed
In use
by my friend and me
6
What I learned
Rubixo is where I practise the whole job: the thesis, the privacy model, the process and the final call on every AI recommendation.
An AI recommendation stays Proposed until I move it, and a CI gate fails any pull request that touches application code without a brief.