INNOVATION / WHY WE EVOLVE WITH EVERY PROJECT

Innovation, the name of the game.

With every project we ship, we try to become a better version of the studio we were last quarter. We keep ourselves abreast of the tools, technologies and trends that move through this industry — and we put time and capital behind the ones we believe in, well before a client asks us to.

LAB WEEK · NOTHING TO SHIP

An hour at the whiteboard.

Most of our R&D output starts here: arrows, boxes, a question someone has been turning over for a week. Once a quarter the whole studio steps out of client work for five days, and the diagram on the wall is the only deliverable that matters.

Read the essays →
CHAPTER01INTENT

What keeping current looks like, in practice.

Most studios in Hong Kong are paid to deliver last year’s stack on this year’s deadlines. There is little internal pressure to learn anything new before a client asks for it, and the work quietly drifts a year or two behind.

We try to invert that. A fixed share of every working week, and one full week per quarter, is reserved for work the studio pays for itself — internal tools, AI experiments, infrastructure rewrites, essays, conference talks. Each engineer’s ten-percent learning time runs in parallel.

The output ends up in client engagements the next quarter — cheaper, faster, and with fewer surprises than if we had learned on someone else’s clock.

CHAPTER02ACTIVE LABS

Four things we are figuring out.

LAB / 01ACTIVE · SHIPPED IN CLIENT PROJECTS

An AI evaluation harness

A small, opinionated framework for measuring whether a model or prompt change actually improved the output. Built because the off-the-shelf eval frameworks are either too academic or too tied to one vendor.

PythonSQLiteLiteLLMStreamlitMIT-licensed
LAB / 02ACTIVE · INTERNAL

A pattern library for HK SME apps

Reusable building blocks for the patterns that recur across HK retail, F&B, and finance apps: bilingual content management, MTR-friendly nav, HKID validation, Octopus / FPS integrations, member-tier UX.

SwiftUIComposeReactTailwind
LAB / 03ACTIVE · RESEARCH

Small-model retrieval for regulated data

Investigating how far open-weights models (Qwen, Llama, Mistral) can be pushed on dense legal & financial retrieval, running entirely on hardware the client controls. The answer matters for every HK advisory firm we work with.

Qwen 2Llama 3pgvectorBM25 + densevLLM
LAB / 04ACTIVE · INTERNAL

A build-system for embedded squads

The tooling that lets one or two of our engineers sit inside a client’s operation and ship faster than a team of five would: provisioning, observability, internal-admin scaffolding, customer support tooling.

TerraformOpenTelemetryNext.jsPostgres
CHAPTER03PRINCIPLES

What our R&D does not look like.

PRINCIPLE / 01

No moonshots

We don't fund speculative bets on five-year horizons. Every lab project must plausibly show up in a real client engagement within twelve months. If it can't, it isn't a fit for us — that is for venture capital, not for a services studio.

PRINCIPLE / 02

Boring beats novel

A small win on a boring stack (Postgres, a cron job, a clean queue) usually beats a flashy framework that nobody on the team can debug at 3am. Our internal tools are deliberately under-engineered.

PRINCIPLE / 03

Open by default

If a tool we build for ourselves has no commercial value to AccordHK as a product, we open-source it. Not because the world demands it, but because the act of preparing a repo for the public forces us to write the documentation properly.

PRINCIPLE / 04

Writing is the work

An essay shipped is research published. Every senior engineer is expected to publish one piece of long-form writing per year — on what they learned, what they got wrong, or what they think the rest of HK should be doing.

CHAPTER04OPEN SOURCE

Six repos we maintain in public.

accord/evals

LLM eval harness

The small, opinionated framework from Lab 01. Used internally on every AI engagement; published as a starting point for HK SMEs.

MIT · v0.4
accord/forms

Bilingual form kit

A small TypeScript form & validation kit that ships English and Traditional Chinese error messages in lock-step. The validation rules our HK clients keep re-discovering they need.

MIT · v0.6
accord/zh-tokeniser

繁中 search tokeniser

A Traditional-Chinese-first tokeniser tuned for HK product search. Outperforms the default Postgres + jieba pipeline on small HK corpora.

MIT · v0.7
accord/octopus-ux

Payment-flow UX kit

Reference flows for HK payment methods — Octopus, FPS, AliPay HK, WeChat Pay HK — that have actually shipped in production apps, with their gotchas documented.

CC-BY · living doc
accord/cli

Project scaffolder

The CLI we use to spin up a new client project: Next.js + Postgres + auth + a sensible analytics setup, ready to ship in an hour.

MIT · v1.2
accord/style

The studio style guide

How we write code, write English, write Chinese, write commits, write postmortems. Public because new joiners shouldn’t have to wait until day one to see how we work.

CC-BY · living doc
CHAPTER05PARTNERSHIPS

Who we collaborate with.

ACADEMIC

HK universities

Adjunct teaching slots, internships and applied research collaborations with two HK universities — quietly, on terms that respect everyone's IP.

VENDOR

Cloud & model vendors

AWS Advanced Tier, GCP partner, plus formal access programmes with Anthropic, OpenAI and Mistral for client deployments needing reserved capacity.

INDUSTRY

HK SME associations

Pro-bono engineering hours for two HK SME associations, and yearly talks on AI adoption for trade groups in F&B, retail, and logistics.

REGIONAL

East Asia studios

Trusted partner studios in Taipei, Tokyo and Singapore for projects that need local presence beyond Hong Kong. We pass work both ways.

/ / RESEARCH & PARTNERSHIPS

Building something together?

If you’re a researcher, a vendor, an industry association or another studio looking to collaborate — write to us. We’re selective, but we always read.