From our test cards, testing desirability, financial viability, and feasibility, we were able to create our learning cards (as listed in our drive). For our desirability learning card, we tested 6 users by giving them a realistic scenario where they needed to find a cross-team collaborator with relevant experience, comparing their normal workflow against our interactive v0 prototype. All 6 users successfully completed the search and appreciated seeing why each person was recommended, with 5/6 preferring Chilla for outside-network discovery while 1 still favored asking coworkers directly when they already knew the owning team and wanted social context. For the financial viability learning card, we gave 6 users full access to the interactive prototype’s profile querying and calendar optimization workflow, concluding with an in-app advocacy prompt and exit evaluation. While 5/6 users (83%) said they would continue using Chilla for projects involving unfamiliar teams, only 4/6 would actively advocate for company adoption because skeptics questioned whether a standalone discovery tool adds enough value beyond Slack, internal directories, and calendars they already pay for. Finally, for the feasibility learning card, we walked 6 users through a Slack onboarding flow that requested opt-in permissions to share internal work profile data for matching and connect their calendar for automated scheduling. Five out of 6 users approved sharing professional profile details and 3/5 (60%) of those also connected their calendar, whereas hesitant users worried about exposing private meeting information despite disclaimers stating event notes and attendee details wouldn’t be accessed.
From our learning cards, we made a few targeted changes to our prototype. In order to better support users who normally rely on warm introductions, we added degrees of separation (1st, 2nd or 3rd) and mutual contacts to every collaborator card and profile, showing who you both know and how they know the person. The “Ask for an intro” button sends the request through that shared contact, so users can gauge how comfortable they feel reaching out and get a warm introduction instead of a cold message. Also rather than positioning Chilla solely as a people-finder, we expanded the prototype to support the full coordination loop: it takes users from finding the right teammate and understanding why they’re relevant (including any gaps in their experience), to scheduling time, tracking the dependency and automating follow-ups. Confirmed meetings are now blocked on both calendars, so the same slot can’t be double-booked. We also reduced privacy hesitation during Slack onboarding. We rewrote the calendar permission prompt to list exactly what Chilla can and cannot see, and added a preview of the only data it receives, which is free/busy blocks with no titles, attendees or notes. Chilla also enforces this in conversation: if someone asks what’s on a coworker’s calendar, it declines and offers to find a mutual free time instead. Further changes from our own evaluation include making Chilla honest about what it doesn’t know.
For our test cases, we initially failed many of our edge cases. Chilla would often return people who were not relevant to the tasks. The biggest assumption this proved wrong was that showing loosely related people is better than showing no one; in practice, a confident wrong match hurts trust faster than an honest “I don’t know.” So when Chilla is asked for information that does not exist, it now makes it clear that it doesn’t know the answer instead of responding with loosely related people, and offers to search the part of the question it does recognize. We also made sure that once a meeting is confirmed, it is immediately blocked off on both people’s schedules, so the same time slot can no longer be double-booked.
Our additional evals also showed that Chilla struggled with more complex requests, particularly follow-up questions, comparing expertise across teams, and finding people who meet multiple requirements. It also sometimes gave very high match scores to people whose experience was only loosely related to the request. We improved how it handles conversational context so it can answer these requests, and made its recommendations more transparent by showing what each person has direct experience in and flagging any gaps. After these changes, all 9 eval cases pass. Because we tuned against those cases, our next step is to write a fresh set of test cases to check that the fixes generalize.
In terms of our OKRs, we used the concepts from the reading Christina wrote. The reading states this idea of making an objective aspirational, memorable, and qualitative, to capture why the work matters so it motivates the team. We applied this idea directly to our own objective and came up with: “Running over our competitors and making cross-functional collaboration and 1:1 discovery completely frictionless, transforming team coordination into an intuitive, empowering and enjoyable experience.” We want to highlight that we are creating empowered individuals who can work efficiently and find joy in this work. This also translated into our key results, measuring through unambiguous numbers that prove wear goals, as the reading suggests. We made clear objectives aiming at retention, friction reduction, and desirability.
