All work

    Zevian

    Weekly performance scoring that managers trust, live with paying customers

    Managers write the criteria. Zevian scores every weekly report against them and ties each score to a quote from the report. I founded it and directed the build, so every product and design decision below was mine.

    4.4 / 5

    Average usefulness, across 5 manager interviews

    4 of 5

    Managers willing to try it at the price I tested

    2

    Paying customers, from direct outreach

    Role
    Founder. Product, design, data model, prompts, and sales
    Team
    Solo
    Platforms
    Web app, multi tenant

    Three decisions that shaped the product

    1. 01

      I tested the price with real managers before I built onboarding or marketing.

    2. 02

      Every score points to a quote from the report. A manager can change any score.

    3. 03

      The assistant only answers from the reports it retrieved. If the reports do not cover the question, it says so.

    0 to 1 product strategyMulti tenant system designAI scoring designAI assistant designData model designPrompt designDesign to engineering handoff

    One team score, one reporting bar, and the numbers each agent hit against target.

    Chapter one

    Problem, strategy, and process

    CRMs show deals moving. HR tools run reviews twice a year. Nothing tracked the weeks in between. Managers scored from memory, employees got vague feedback, and a bad hire went unnoticed until month four.

    Two discovery interviews, four follow ups, and a survey with 16 responses. The problem was worst on teams of 15 or more.

    Strategy

    One question mattered more than the rest: would a manager pay to have their team scored. Until that had an answer, anything I spent time on was a guess. So I tested price first, shipped a plain interface early, and spent design time only where managers got stuck.

    Validate demand first

    I tested $8 per user per month during discovery, before onboarding or marketing. Four of five managers said they would try it.

    Keep the interface cheap to change

    I shipped rough screens on purpose. Every rebuild came from watching someone use the product, not from polishing alone.

    Almost nothing in the product is a first draft. The loop was the same every time: ship it, watch someone use it, rebuild it.

    Rebuild indexThe full saw / changed list, kept out of the main reading flowShow
    AreaWhat I sawWhat I changed
    WeightingManagers had to make their percentages total 100 themselvesImportance levels
    CriteriaJudgment criteria only. A customer wanted his team's reported numbers next to the scoresNumber criteria
    AI contextThree separate fields. Managers asked which one changed the scoreProject Memory
    DashboardStatus flags on every row and four KPI cards that looked the same. Nothing said what to check firstOne main number
    Employee dashboardOne Overview tab, a color per KPI, reasoning one criterion at a time. Seeing if a person improved took clickingFive tabs
    BadgesImproving and Aligned on every row, so the badge on the person who needed attention did not stand outExceptions only
    AssistantA call center manager pasted his team's numbers into ChatGPT and acted on the answerScoped assistant

    Each change links to the section that covers it

    Chapter two

    Manager workflow

    A KPI is a goal, written instructions, and a set of weighted criteria: a scoring rubric. Every score in the product is measured against that definition.

    Weighting

    Problem

    Managers typed a percentage for each criterion. Making the numbers reach 100 took more thought than the work itself.

    Decision

    Split KPI creation into two steps: define the KPI first, then add the scoring rubric. Presets and the AI builder help managers start from a role instead of a blank form.

    Why

    It asks managers to decide one thing at a time: what success means, then how each report should be scored. Importance levels still set the weight automatically.

    Tradeoff

    Exact control is gone. Nobody has asked for it back.

    Before and after: one long form became a two-step flow with presets, AI-assisted setup, and criteria added after the KPI details.

    Number criteria

    Problem

    It shipped with judgment criteria only. A customer wanted the numbers his team already reported shown next to the scores. He wanted to compare the two.

    Decision

    Number criteria carry a target, a unit, and a direction. The system checks the number against the target. The AI does not score them.

    Why

    He was right. A manager who reads one number, or the same number across a few weeks, already understands it. An AI in the middle adds a step and a reason to doubt the answer.

    Tradeoff

    Two criterion types to learn, and someone has to keep the targets up to date.

    Which type fits depends on the team. Sales floors and call centers run on numbers and on catching a pattern early. Work that is harder to count stays on judgment criteria. Most teams use both.

    Project Memory

    Problem

    Three separate fields: scoring instructions, project description, and knowledge base. Managers kept asking which one changed the score.

    Decision

    Merge all three into one plain text box. Memory only updates when the manager presses Update.

    Why

    The three fields came from how the system was built. A manager does not think in those parts. Manual updates mean every change to scoring is on purpose.

    Tradeoff

    Less structure for the model. In exchange, the manager knows what they are editing.

    One box for everything the AI reads.

    Score guard design

    Problem

    An AI that rewards polished writing teaches employees to polish reports and stop there. The product would create the problem it was sold to fix.

    Decision

    Rules that treat writing as a claim that needs proof.

    Why

    Every score has to point at something the person wrote. That is what makes it hold up in a review.

    Tradeoff

    Someone who writes briefly but did good work scores lower than they deserve until they add details.

    Anti-gaming rulesHow the score guards against polished or repeated reportsShow
    • Pattern detection — templated reports: the same report submitted each week with different wording.
    • Prompt detection — entering an instruction as-is with a few changes.
    • Vague claims — scored less.

    What is in a report

    The two criterion types are shown differently. The pattern flag looks across several weeks.

    Pattern flagged, numbers checked against target, and judgment scores anchored to report evidence.

    Manager dashboard

    Problem

    Status flags on every employee row, then the same numbers again across four near identical cards. Nothing said what to check first.

    Decision

    Move each trend into the card it belongs to. One main number with its trend line inside it, and every report status in a single bar.

    Why

    A manager opens this page with one question: is anything wrong this week. One number and one bar answer it. The team list below shows which person.

    Tradeoff

    Detail moved a click away, which is why I built the scoped assistant below.

    Before

    After

    Manager override

    Problem

    The AI will be wrong sometimes. A manager who cannot change a score will not trust it with people decisions. A person has to decide.

    Decision

    Managers can change any score. A reason is required, and both numbers stay in the record.

    Why

    After eight adjusted reports, the pattern says as much about the manager as the employee: Likely Inflated, AI Aligned, or Underscoring.

    Tradeoff

    Every override takes extra effort. That effort is what makes the record worth having.

    No reason, no override.

    Scoped assistant

    Problem

    A call center manager pulled his team's numbers out of the system and pasted them into ChatGPT or Claude. He asked which of his people were falling behind. It answered with confidence, it was sometimes wrong, and he made decisions about people based on it.

    Decision

    Build the assistant into the product and scope it to one source: the reports the team submitted, filtered by group and date range.

    Why

    A general model answers anything, which is what went wrong. Retrieval augmented generation limits this one to reports that exist, and it names them under every reply.

    Tradeoff

    It refuses far more questions than a general model would. Refusing is better than guessing about a person.

    Suggested questions it can answer.

    The search step is shown as it happens.

    Annotated: the scope filter, the reply, and the reports it drew from.

    An answer from retrieval augmented generation needs three parts:

    • What it was allowed to read
    • The answer
    • The sources, naming every report the answer came from

    Chapter three

    Employee trust and system

    Problem

    A scoring product only works if the people being scored accept it. If it feels like being watched, they write for the AI or stop caring.

    Decision

    Employees see the same numbers the manager sees. They see the criteria and weights before they start writing. A draft can be scored and revised before it is submitted.

    Why

    Scoring a draft lets someone fix the report against the criteria before anyone else reads it. That is why employees started using it.

    Tradeoff

    Showing the criteria means people write to them. The anti gaming rules mean the only way to score well is to do the work. The reason behind a manager's change stays private, so nobody can work out the gap between AI and manager.

    The criteria and weights sit above the writing box.

    Employee dashboard

    Problem

    One Overview tab held everything, and every KPI had its own color. The reasoning for a low score opened one criterion at a time.

    Decision

    Dedicated tabs, one shared KPI color, and a six-period table with color only where attention is needed.

    Why

    The table answers the main question: is this person improving. No clicking.

    Tradeoff

    Six periods is a guess. Long enough to show a trend, short enough to fit on screen.

    Annotated: the old overview tab next to the rebuilt employee dashboard.

    Badges

    Problem

    Every row carried the same two green badges, Improving and Aligned. The badge on the person with low scores did not stand out. Improving was measured against an all time low, so almost everyone qualified.

    Decision

    Reserve badges for exceptions, and measure Improving against recent scores. A filled badge is a fact about the person, an outline badge is a trend from their scores.

    Why

    The color system was already set: green good, amber caution, red critical. My first instinct was to give every badge its own color, which would have broken that. A badge is only readable if it is rare.

    Tradeoff

    A row with no badge says nothing, so a manager has to trust that no badge means no problem.

    Before

    After

    Shipped: 28 people, 2 badges.

    System

    The scoring model sits on a multi tenant data structure. KPIs are stored as weighted criteria, and each one is scored a criterion at a time against the manager's own instructions.

    The data model behind scoringMulti tenant structure and the scoring schemaShow

    Built for many companies at once

    • Every employee, project, KPI, and report belongs to one organization
    • A three layer role based access structure decides who sees which data. Managers at different levels only see their own teams

    KPIs stored as weighted criteria

    • Each score is saved one criterion at a time, with its own instructions, grouped under a KPI
    • That is what lets one skill be tracked over time, for the company and for each person

    AI that scores evidence

    • Every score is tied to a quote from the report and to the manager's own instructions
    • Nothing scores well on writing quality alone

    Organization → Employees → Projects → KPIs → Reports → Criterion Scores

    Outcomes

    What shipped

    4.4 / 5

    Average usefulness, across 5 manager interviews

    4 of 5

    Would try it at the price I tested

    2

    Paying customers, from direct outreach

    Live

    Scoring, overrides, onboarding, notifications

    The gap is usage data. PostHog tracking is going in now. Funnel and retention numbers will sit here once there are enough customers for them to mean anything.