AI-assisted decision workspace

What getsapproved iswhat runs.

Turn business requests into clear requirements, informed decisions, and controlled action.

Portfolio prototype · synthetic data

From uncertainty to action

  1. Understand
  2. Evaluate
  3. Approve
  4. Execute

01 · The problem

Requests arrive unfinished. Decisions shouldn't.

Briefs come as emails, PDFs and chat messages. Targets are clear, but the details that decide a deal are missing or contradict each other. An AI can read faster — it can also guess.

Karsa reads the brief, marks every gap, and never fills one with a guess.

brief — Aster Labs (synthetic)

Target: 3,000 qualified leads

Budget: Rp150,000,000 for the full campaign

Timeline: 2 months from kickoff

Geography: to be determined

We have not yet agreed what counts as a qualified lead.

Finance note: the vendor cap is Rp120 juta.

  • Confirmed target, budget, timeline — each with its exact quote
  • Missing geography — the brief says it is still open
  • Missing lead definition — asked, not guessed
  • Conflicting budget — both values shown, a person decides

02 · How it works

Five steps. A person signs off on every one that matters.

01

Evidence, not guesses

Each value points to the sentence it came from. No source? It is marked missing, not guessed.

02

Rules do the math

Cost and margin come from the rate card, not from the AI.

3,000 × Rp35,000+ Rp12,000,000= Rp117,000,000margin 22%

03

Approve the exact snapshot

Approval locks to one hashed version. High-risk deals need two different people.

04

Execute only what was approved

If an outcome is unclear, retry stays off until someone checks the target system.

05

Measure honestly

Scored on 36 cases it never saw: 1 passed. The release gate blocks it.

03 · Principles

AI where it helps.
People where it matters.

A source, or visibly missing

Values carry their quote and position. Answers from clients need a reference you can check.

Approving is not executing

Two separate acts, recorded separately. Nothing reaches the CRM or tracker on approval alone.

Failures stay visible

Incidents, manual resolution and cancellation with a compensation list. Nothing is deleted quietly.

Honest measurement

Held-out evaluation, sample sizes, and a causation caveat on every impact chart.

04 · What is proven

Tested by breaking its own rules.

Backend tests
51
Browser journeys
16
Pages scanned for WCAG 2.2 AA
12
Prices equal to hand-computed values
60/60

Not yet: a live Claude evaluation, SSO, real CRM connectors, OCR, a hosted demo, or real users.

Prototype on synthetic data · no client or business result is claimed

05 · Built to be used

Decide at your desk. Approve from your phone.

Analysts work in a dense desktop workspace; approvers get a calm mobile review with the numbers, the changes and the evidence one tap away.

Mobile approval screen for a high-risk package with two approver slots and approve, reject and request-changes buttons