←All work

Case study · AI tooling

Hyoka.A code reviewer that never gets tired.

Hyoka (評価, “evaluation”) reviews every pull request with an LLM and leaves inline suggestions the author can apply in one click. I took an early code-review prototype and rebuilt it into a multi-tenant product that whole teams can share.

Role
Lead developer, end to end
Where
dxtr Asia
When
2026
Platforms
GitHub · GitLab
評

hyoka-bot suggested a change · src/billing/invoice.ts

Medium total can be undefined when the invoice has no line items, and toFixed will throw.

Suggested change
−return formatIDR(invoice.total.toFixed(0));
+return formatIDR((invoice.total ?? 0).toFixed(0));
Illustrative exampleCommit suggestion

How it works

From “PR opened” to a suggestion in six steps.

Tap any step, or let it play. The key decision: the API never waits on the LLM. Webhooks get an answer in milliseconds, and the slow work happens on a queue.

  1. PR
  2. API
  3. Queue
  4. Worker
  5. LLM
  6. Done

Step 1 / 6

A developer opens a pull request

GitHub or GitLab fires a webhook at Hyoka. Each team (an entity) has its own webhook secret, platform token and review rules.

hyoka · logs (illustrative)
  • › POST /webhooks/github pull_request.opened #482

The hard part

An LLM that only sees the diff makes things up.

Most bad AI reviews come from missing context: the model can't see the helper you imported, so it tells you to write one. Hyoka lets each project choose how much surrounding code to send. Try the four modes:

  • src/billing/invoice.tschanged lines
  • src/lib/money.tsimported
  • src/billing/types.tsimported
  • src/billing/invoice.test.tsrelated
  • src/billing/tax.tsrelated
  • src/app/routes.tsnot sent
5of 6 files sent to the model

Adds related unchanged code outside the diff window, like the tests and siblings of the changed file. The best balance for most teams.

File names are illustrative.

What I built

01

Multi-tenant from day one

One deployment serves many teams. Each entity gets isolated config and data: its own webhook secrets, platform tokens, LLM keys and review rules.

02

Bring your own model

OpenAI, Anthropic Claude, Google Gemini and DeepSeek through the Vercel AI SDK, with a default per entity and an override per project.

03

Secrets that stay secret

Tokens and API keys are encrypted at rest with AES-256-GCM, using a key derived per entity with HKDF-SHA256.

04

Fewer false positives

Extra file context outside the diff, plus a deterministic filter that drops self-contradicting “missing import” and “missing null-guard” findings.

05

Suggestions you can apply

Line and role validation rejects broken suggestions before they reach the PR, so what's posted applies in one click.

06

A dashboard for the bill

Token cost, request counts and model breakdown by month, quarter or year, so teams know what reviews cost.

Stack

API

Node.js 24

Fastify 5

JWT auth

Jobs

BullMQ

Redis

Data

PostgreSQL

Prisma

AI

Vercel AI SDK

4 providers

Dashboard

React 19

Vite

Mantine

Ops

Turborepo

pnpm

Docker Compose

GitLab CI