AI systems · Co-founder, Plus The Site

AI systems built to be inspected. Every claim on this page ships with evidence.

I’m Rifqi Sigwan Nugraha, a developer in Jambi, Indonesia and co-founder of Plus The Site, a software agency. I build inspectable AI products across Next.js, Supabase, durable workflows, Python, and Flutter, and I follow failures through traces and data boundaries until the real cause is clear.

View flagship build GitHub Download CV

Available remotely

CareCanvas: 19 / 19 tests

NALAR: 24 / 24 tests

Research: 1st author · IEEE

Based in: Jambi · UTC+7
Focus: AI systems
Stack: Next.js · Python
Work: Remote-ready
Status: Open to roles

Best-fit problems

Inspectable AI pipelines
Backend architecture
System integration
Systematic QA
Agentic workflows

Selected work

Two systems you can open and run.

CareCanvas is a human-gated AI illustration pipeline. NALAR is a role-based operations prototype. Both ship with a public URL and a public repository, so nothing here rests on my description of it.

Each case study separates three things: the working product, the engineering evidence behind it, and the boundary of what has been verified.

Flagship · public deterministic demo

01 / 02

Human-gated AI illustration pipeline

CareCanvas

An AI prototype that turns a visual brief into a child-appropriate illustration through bounded agents, an explicit human approval gate, durable execution, and a trace a reviewer can debug.

Run public demo Inspect source

19 / 19: automated tests
2: bounded attempts
96: QA drill score
0: external provider calls

CareCanvas public demo showing its AI illustration workflow and review interface — Live interface · deterministic safe-demo mode

01
BriefStructured intent
02
SafetyStop before spend
03
ApproveHuman gate
04
GenerateLocal or Flux adapter
05
Visual QAOne bounded retry

Implemented evidence

The boundary work is visible.

Explicit state transitions, Inngest approval waits, Claude and Flux adapters, owner-scoped Supabase RLS, raw-body Ed25519 webhook verification, and auditable traces.

Honest release boundary

The demo proves orchestration.

The public demo uses deterministic local adapters and makes zero external provider calls. Anthropic Claude, fal.ai Flux, Inngest, and Supabase live paths remain environment-gated.

Next.js 16
Claude adapter
fal.ai Flux
Supabase
Inngest
TypeScript

Live portfolio prototype

02 / 02

Role-based decision support

NALAR

A cooperative operations prototype spanning sales, POS, customers, receipt verification, national mapping, and optional grounded AI assistance. The core workflow still works when providers do not.

Open live app Inspect source

24 / 24: tests passing
Strict: TypeScript
Safe: degradation

Disclosure: NALAR uses synthetic portfolio data. It is not presented as a production cooperative or a claim of live users.

Server-owned access

Supabase service-role calls remain behind server APIs; elevated credentials never reach public clients.

Optional AI

Grounded Gemini responses use application context and deterministic fallbacks when the provider is unavailable.

Verified paths

Tests cover forecasting, tamper detection, fail-closed APIs, rate limits, and safe degradation.

Next.js 16
React 19
Supabase
PostgreSQL
Gemini
Vitest

Recognition

Evidence beyond the codebase.

Awards, scholarships, and credentials that back the work above. Each one verifiable.

Co-founder · 2025–present

Plus The Site

Co-founded a software agency: end-to-end platform delivery, team leadership across design, development, and marketing, plus selection for the SIAP 2025 acceleration program by Kementerian Ekonomi Kreatif.

Scholarship · 2024–2026

Bank Indonesia Scholarship

Full academic scholarship for excellence and leadership potential through the final two years at Telkom University.

Research · IEEE Xplore

First-author publication

Synthetic IoT data generation study indexed internationally as first author.

IP rights · Kemenkumham RI

Registered software copyright

“Tel-U Great Planner Academic”, Reg. No. 000925616, registered 2025.

Government program · 2025

SIAP acceleration participant

Selected for Kemenekraf’s national acceleration program for creative-economy ventures.

National finalist · Telkom Indonesia

Innovillage Top 163

IoT flood early-warning system; national finalist two years running (2023 & 2024).

Security research · HackerOne

TikTok vulnerability report

Client-side audio-restoration flaw found independently and disclosed responsibly.

Olympiad · Astronomy

OSN provincial finalist

National Science Olympiad finalist at provincial level.

Professional certification

BNSP digital marketing

Nationally certified competency in digital marketing practice (2024).

Facilitator · Dicoding Indonesia

Google Skills Arcade 2026

Leading a cohort through Google Cloud learning paths (Jul–Sep 2026); Gen AI, LLM, and Responsible AI credentials on Google Skills.

Working system

How a change gets verified before I call it done.

The same five steps produced everything on this page: the test counts in the case studies and the smaller systems below.

Claude Code is in my daily workflow, but traces, diffs, tests, builds, security checks, and human review decide whether the work is done.

1
Observe
Reproduce the failure and collect the smallest useful trace, log, or test.
2
Decompose
Split independent analysis, frontend, backend, data, and QA work where parallelism helps.
3
Implement
Keep ownership boundaries explicit and change only what the supported hypothesis requires.
4
Challenge
Test permissions, error paths, security assumptions, and regression risk.
5
Verify
Run tests and builds, inspect the user path, and record remaining uncertainty.

Human integration gate

One accountable owner.

Specialist agents can investigate separate questions. I reconcile their output against the repository and own the final decision. Conflicts trigger more evidence.

01Scoped roles
02Acceptance criteria
03Dependency-aware sequencing
04QA before “done”

Claude Code
Next.js
Supabase
Python
Flutter
GitHub Actions

Supporting evidence

Six more systems with the same verification habit.

Smaller than the two flagship builds, and each one carries evidence you can open: a peer-reviewed paper, public repositories, and running apps.

Where a project has a limit, a retired deployment or credit shared with a team, the card says so.

First-author research · IEEE Xplore

Synthetic IoT data for smart agriculture

Comparative analysis of Gaussian mixture models and Monte Carlo simulation for synthetic data generation in durian-cultivation smart-farming IoT systems.

Monte Carlo
GMM
Smart farming
IoT data

View IEEE record

Public repository · 24 tests

Shariah Trading Assistant

A role-scoped multi-agent research and analysis system with evidence handling and explicit QA gates.

Multi-agent
Orchestration
QA gates

Inspect source

Team build · project lead

Protein RADAR

A Next.js and PostgreSQL dashboard integrating more than twenty public data sources, authentication, RBAC, and audit logging. My contribution was within a documented team.

Next.js
Prisma
PostgreSQL
RBAC

Open live app

Production IoT · Desa Mukai Tengah, Kerinci

Flood early-warning system

IoT water-clarity and flood monitoring platform (Tim Rewana) serving roughly 300 households, with sensor integration, solar power, and a real-time web dashboard.

C++
IoT
Embedded
PHP

Backend engineering · computer vision

AI surveillance dashboard

Backend services, data pipelines, and RESTful APIs for real-time video analytics, intelligent alert management, and uptime monitoring, integrating 10+ computer vision models over WebSocket.

Python
REST API
WebSocket
Data pipelines

Computer vision · archived deployment

Qoffea

A coffee-bean assessment workflow with Flask, YOLO inference, confidence filtering, non-maximum suppression, API tests, and container configuration.

The former deployment is retired; this links to source evidence only.

Python
Flask
YOLO
Docker