Kaccharts Portfolio

Operations

MainStory · 2025Product Design LeadShipped

How a Retool Sheet Became an AI System

QC Audit · replacing 3 hours of manual quality review with 15 minutes of supervised AI scoring.

MainStory had 10 daycare centers and a quality problem. Three operations people spent ~3 hours per center every morning reviewing CCTV, validating timeline posts, and chasing missing meal entries. Then an incident happened. We couldn't scale the manual process — not safely, not affordably.

I designed a three-layer audit logic: deterministic checks for timestamps and required fields, pattern matching against the menu config, and KinderGPT validation for fuzzy meaning ("kerupuk" is a valid food; "ok" is not a valid caption). Scoring was progressive, not binary — CMs see exactly why each audit landed where it did. Trust is built in transparent breakdowns.

100%Automation achieved

98%Time reduction (3h → 15m)

98%Compliance rate (target: 95%)

Read the full case study →

Workforce · Back-end

MainStory · 2025Product Design LeadShipped (3 of 4)

The Caregiver OS

Designing how 200+ caregivers get paid, rated, and trusted — across four interlocking products.

Before this work, MainStory's compensation lived in a Google Sheet, partially in someone's head, and partially in a Retool tool that didn't know the rules. Every payroll day produced the same monthly conflict cycle: "kenapa gajiku segini?" and CMs without the data to answer.

I treated this not as a payroll fix but as a system to redesign — four interlocking products: the canonical payroll spec, a variable income layer (rating + tipping), an overtime reconciliation flow that closes the trust gap between parent payment and caregiver payout, and a Talenta-style payslip transparency screen where every rupiah is explainable. The math is always visible, always traceable.

35.5%Tipping adoption (target: 10–20%)

88%Overtime pre-paid by parents (target: 80%)

200+Caregivers on the new system

Read the full case study →

Growth

MainStory · 2025Product Design LeadShipped

Monthly Wrapped

Turning an internal monthly report into a viral acquisition channel.

Center Managers were spending hours per child every month assembling slide decks for parent calls. Parents loved them — and were already screenshotting and forwarding them to friends-with-toddlers. The deck wasn't the product. The screenshot was.

I designed a 9-card vertical artifact (Spotify Wrapped meets daycare) auto-generated by KinderGPT — same AI infrastructure as QC Audit, redeployed. Four prompt templates, a quality-check pipeline, CM-as-feedback-loop, two-channel distribution. Every share is a free acquisition touchpoint. The internal artifact became a growth channel.

Monthly → QuarterlyCM call cadence

Hours → MinutesPer-child production time

2×AI systems on shared KinderGPT infra

Read the full case study →

Workforce · Front-end

MainStory · 2025Product Design LeadShipped

Owning the Funnel That Fed the Workforce

A 0→1 caregiver hiring platform that cratered for 30 days, then 5×'d.

MainStory was supply-constrained. Glints and JobStreet produced a flat ~220 caregiver leads per month. We couldn't scale centers without our own funnel. I built a landing page (Stage 1) and a full Lovable hiring platform (Stage 2), redirecting acquisition spend away from third-party boards.

Lead inflow dropped to 7% of baseline in the first month. Several stakeholders wanted to revert. I held the line — defining the leading indicators in advance, distinguishing the expected dip from real failure. By month 3, the platform was producing 4× baseline. By month 6, more than 5×. Owning your funnel beats renting it. But you have to budget for the dip.

+435%Lead inflow at 6 months (target: 30%)

−93%At 30 days (the dip)

+322%At 90 days (the recovery)

Read the full case study →

Orchestration

MainStory · 2025Product Design LeadShipped

The Notification Board

Replacing seven scattered notification channels with a single source of truth.

One in eight monthly billings was being missed or delayed at MainStory — not because notifications didn't exist, but because they existed in seven places at once and no place was authoritative. The fix wasn't an eighth channel. It was a board: persistent, queryable, with two authoring paths (automated triggers and Ops-configured announcements) and nine deterministic click-through routes.

Every cycle, fewer corner cases got lost. The improvement compounded — not on day one, but every cycle after. By month 6, missed billing was 84% lower than baseline. Notification Board became infrastructure other products read from.

−84%Missed billing at 6 months

13.27% → 2.11%Baseline → 6 months

−55%By month 3 (compounding)

Identity / Data Architecture

MainStory · 2026Product Design LeadIn Progress

Migrating Without Migrating

Three platforms, three databases, three statuses for the same person — unified into one account architecture.

A mitra asuh at MainStory was simultaneously three different people. "Passed" in the hiring Supabase, "active" in the training Supabase, nonexistent in the kiwi-api MySQL until a fragile edge function copied her over. The default fix is a bulk migration. Bulk migrations are where companies break.

I designed a dual-branch user-driven migration: legacy users self-migrate via a "Perbarui Account" flow on next login; new users skip Supabase entirely. Both paths converge on the same BE schema. No migration weekend. No cutover moment. Plus a unified status taxonomy (MECE, with structured notes), eligibility gates instead of separate accounts, and link-by-identifier handling for multi-portal applicants.

3 → 1Databases

0Migration weekends required

30–42dTotal engineer-day estimate

Read the full case study →

AI Product

Sirka · 2022UX Researcher → DesignerShipped

From Manual Logging to AI Recognition

Freeing nutrition coaches from calorie math — three years before "AI-augmented design" was a phrase.

Sirka's coaches were each managing 60 clients, with a target of 105. The Meal Log feature meant to make them productive was actually consuming their time — 80% of entries were manual, requiring coach calorie math. Research revealed the unlock: users weren't choosing manual because they liked it. They wanted instant feedback, and the database wasn't fast enough.

I designed a photo-based AI recognition flow (ChatGPT-era LLM parsing + Sirka's existing food database + FatSecret/NutriSurvey datasets). The flow collapsed from "search → select → adjust → wait for review" to "snap, describe, done." Coaches stopped doing arithmetic and started doing coaching.

−60%Coach workload (4–5h → <2h/day)

+75%Feature retention (40% → 70%)

30m → <1mPer-meal logging time

Read the full case study →

I design operating systems,
not interfaces.

Seven systems, six layers, one thesis.

How a Retool Sheet Became an AI System

The Caregiver OS

Monthly Wrapped

Owning the Funnel That Fed the Workforce

The Notification Board

Migrating Without Migrating

From Manual Logging to AI Recognition

Hi, I'm Aqis.

Experience

Education

Awards

Skills

Languages

Earlier projects across healthcare, telco, fintech, and consumer.

Habit Challenge Program Revamp

CM Monitoring Log

Supply & Demand Automation

Share as Frame

Finance Automation (Jurnal)

Builder.io CMS Migration

Improving the User Spend Journey

Onboarding Conversion Research

Aloshop PSEF

Nussa Website

Grosenia B2B Marketplace

Ecohero

If you're hiring for someone who designs operating systems, I'd love to talk.