- A new admin console opened, with one screen you can look anything up from. Until now the team's own tools lived in an older internal console; a fresh one now stands beside it and starts taking over. Its Explorer is one place to begin: search for a person or a project, click a person to see their record and every project they own, click a project to see the whole conversation with Mira plus the job description, the files, the business details, the search preferences, the talents and the raw record, all as tabs on one page. Who can see what stays deliberately separated, so someone who only handles support questions still cannot open the conversations.
- Raised flags now have a proper worklist. When the system flags something for a human, those flags sit in their own list to be worked through and cleared. It is kept as a to-do list on purpose, with no trends or scores mixed in, and when the list is longer than what fits, it says so rather than quietly showing a slice and letting it read as everything. The links inside the alerts we post to chat now open the console that actually exists, instead of a page nobody publishes.
- A talent management dashboard, built on the analytics warehouse. A first real view of how our talent side is doing: a strip of headline figures and a full, filterable, sortable table of sellers. It reads from the curated analytics data rather than the live product database, so browsing it can never slow the product down, and it refreshes on its own every hour. Unknown is never dressed up as zero: a seller nobody has approached yet shows a dash, not 0%, an unmeasured response time shows a dash, not "0m", and missing information is shown as its own "Unknown" group you can filter to. The freshness of the numbers, and how much of the data is actually covered, sit permanently above the table, and if a refresh fails the page says plainly that it is stale instead of pretending.
- The scout stops ranking a person who misses what you asked for above one who does not. This kept coming back: someone brilliant at the craft but in the wrong country, or missing a language you had explicitly asked for, still landed above people who met everything, because the miss was only ever written on their card as a note and carried no weight in the ordering. Fit is now judged on two separate things: how good they are at the actual craft, and how much of what you asked for they fall short of. Within the same craft level, whoever misses the least comes first, before any consideration of how strong a seller they are, and something you said in your own words counts far more heavily than something the system merely inferred. When it has to loosen a requirement to fill the list, it loosens its own guesses first and never your words.
- The card now shows the package that actually does your job. Sellers list several tiers, and until now the tier shown was chosen on price alone, so a card could quote a small starter package for a job that clearly needed the bigger one. The scout now reads what each tier actually delivers, the number of screens or pages, the source files, the revisions, picks the tier that genuinely covers your job within your budget, and judges the price on that tier. So the price on the card is the price of the work you actually need.
- The search stopped quietly killing itself. A check of a full day of searches found something bad: whenever the scout tried to widen its hunt by telling the search "skip the people I already found", that search failed outright and returned nothing. It was naming those people one way while the search engine expected another. The names are now translated before the search goes out, so the scout's own way of digging deeper works instead of shutting a whole search lane down.
- A dead website no longer turns into your "market". A tester's website on file was a parked, for-sale domain. Our research spotted that correctly, then carried on researching domain parking as a subject, and the articles it read about registrars and spam policy were shown back to that client as their competitive landscape on a home-renovation job. Research now has to state plainly what the address turned out to be, live, parked or unreachable, and when there is no real site behind it, it writes the single honest finding saying so and stops. Nothing from a dead address reaches the client's page, and Mira is told to ask for the right address instead of the note being swallowed.
- Two things Mira was getting wrong in her questions. A tester could only click today in the deadline picker, because Mira was setting both the earliest and the latest allowed day to the same date. The floor is real, a deadline is in the future, but there was never a reason for a ceiling, so it is gone for good. And when someone picked three of four options and deliberately left one out, Mira's very next message asked about the one they had left out. Leaving an option unpicked now reads as a clear "no".
- A talent whose offer is not going forward is told so, plainly. After the back-and-forth on price reaches its limit, the closing message used to just say the offer was recorded and someone would be in touch, never actually answering whether it had gone anywhere. It now thanks them, says clearly that this one is not moving forward because expectations differed, and tells them they are welcome for future work.
- A big reviewed-wording pass through the client app. The product now says "job description" rather than "brief" throughout, and "Mira" rather than "Concierge"; the account menu says "Your projects"; the button on a talent card reads "Message" plus their name; the "job description ready" band, the hint popup and the brand-style intro all took the reviewed wording; "by Fiverr" is back in the header; and the delete warning on your profile now names exactly what goes, your Mira conversation history including briefs, files and shortlists.
- Two small interaction fixes. On the About You page, tabbing between the basic fields used to drop you out of the form entirely while a field was saving, so the whole section briefly became unreachable by keyboard; it now stays put. And the little "Added to brief" and "Mira will remember" notes in the chat no longer look clickable, with no button styling and no tooltip, because they report what just happened rather than offering you a next step.
- Quiet groundwork. A read-only health check we can run on demand fires real searches at the live Fiverr search service and reports whether it is up and returning complete people, without touching or changing anything. The production reporting gateway went in, the emergency stop control moved into the system-admin area where it belongs, signing out of the admin side now properly clears the session, and a stack of deployment plumbing was straightened out so the two admin sites publish to the right places.
The biggest day in weeks, 54 commits across the whole crew. A new admin console opened with an Explorer that puts a person, their projects and a whole conversation one click apart, a proper worklist for raised flags, and a talent management dashboard built on the analytics warehouse that never touches the live product database and refuses to show unknown as zero. On the matching side the scout stopped ranking a must-missing person above a compliant one, started showing the package tier that actually does your job, and had a real bug found and fixed where every attempt to widen a search was silently failing. Research stopped turning a dead, parked website into a client's competitive landscape; Mira's deadline picker lost its bogus ceiling and a de-selected option now reads as a firm no; a talent whose offer is not going forward is told so plainly; and the client app took another reviewed-wording pass, now saying job description rather than brief throughout.