Yunshu Zhiliang · Tech Stack

AI delivery, built as
a pipeline you can verify.

Not another chatbot. We put data sources, tools, agents and publishing into one foundation and run it on your own machines: it takes work by itself, reports on schedule and improves after every review. Every number on this page can be checked by you.

90Production deployments
measured, verified 2026-11-05

500+Registered tools
across 12 MCP gateway servers

37 millionCourt documents
32,296 regulations indexed

See the evidence Talk about your case

The questions we hear most

Before parameters, the three things you actually worry about.

Once you take delivery, three things are most likely to go wrong: where your data lives, whether the conclusions are usable, and whether it degrades over time. Here is each answer.

Will my data be moved somewhere else?

One command installs it on your own machines; data lands in a local SQLite file and is not bound to a cloud account. The full stack can be delivered into your data centre and core capability keeps working offline. You can verify before and after signing.

Can I trust the conclusions it gives me?

Every conclusion carries a source and citation chain that can be recomputed and traced. Anything whose origin cannot be resolved is deleted rather than invented. Contract outputs also pass a three-reviewer pipeline: legal, commercial and completeness.

Will it get worse after three months?

Every run leaves a trace, a judge reviews it, and improvements are written back into skills and the ontology: the team gets stronger with use, not driftier. The self-healing evaluation loop is published in our AI Knowledge library.

One path

From source to usable data in one path, without stitching seven systems together.

Incomplete, inconsistent and non-retained are the three roots of unusable data. This path is ordered against exactly those three.

Data sourcesInstalled as a package

Each source is an independent, auditable data package; public government data, statistics and industry collection all arrive in the same form.

One ontologyOne definition

All sources land in the same semantic schema, so indicators, regions and periods line up across sources.

MCP gateway500+ tools

Capabilities become tools an agent can call: one question returns real data rows without writing code.

Agent clusterAlways on

Organised by role skeleton and handover rules, running 24/7, escalating exceptions to people.

OutputsDashboard · report · API

Three delivery forms; data stays on your side and every refresh can be recomputed.

The same agents, skills and MCP supply can also be loaded into mainstream AI coding tools such as Claude Code and Cursor, so you are not locked into our platform.

Delivery

How a project is delivered: six stages, with the reuse ratio written down.

The percentage on each stage is the portion that reuses the foundation instead of being rebuilt. It rises toward the end because the fixed parts were completed earlier.

Solution design90 capability specifications are called directly; no solution rewrite

70%

Full-stack buildOne contract feeds four clients: web, mini program, macOS and Windows

80%

Engineering qualityTests and CI guards block problems at commit; nothing reaches the main branch

90%

Packaging and releaseThe same CI pipeline outputs both macOS and Windows installers

100%

Launch and operationsThree steps bring up the full stack; data stays with the customer

100%

Knowledge transferSpecifications, documents and code live in one repository

90%

85%Weighted overall reusability across the six stages (15 / 40 / 15 / 10 / 10 / 10 hours). Note: weights and ratios are team estimates, not third-party audited.

Capacity

Ten thousand agents on one machine, without costs rising with the count.

Online scale is amortised by density; real concurrency is sized by hot units. These are the measured capacity tiers.

≈0Idle session on disk
memory per unit

1.6–1.9 GBWarm co-hosting
10 agents per pod

130–190 MBHot unit
single agent

1536 MiPer-cell budget
hard cap

The conclusion of tiered hosting: everything is online, but not everything is resident. Idle sessions sit on disk at near-zero cost, warm co-hosting amortises a single process, and real concurrency scales elastically by worker slots and runner shards.

Evidence

Open it now, without waiting for a scheduled demo.

Four entries, three with no login. If a number looks questionable, go and check it yourself.

Open sandboxTalk to the agent team directly, no sign-up needed demo, no login Natural-language data queryAsk "China GDP 2023" and get real rows with unit, frequency and source public query box Data asset catalogue57,000+ indicators searchable without login, with quality profiles catalogue, no login Open-source repositories21 repositories, 3 published npm packages, MIT licence GitHub

See it first, then talk about working together.

We can run one real dataset of yours through the whole path in half an hour: ingestion, indicators, dashboard. Whatever does not work, we will tell you on the spot.

Try the open sandbox Book a live run

How these numbers are defined: 90 production deployments, 75 unique images, MCP gateway with 12 servers / 500+ tools, 57,000 unified indicators registered, open data ontology with 452 concepts / 7 sources / 53 tools, 37 million court documents, 32,296 regulations indexed. Capacity tiers come from the published hosting model; the six-stage reuse ratios and the 85% weighting are team estimates. All public numbers were verified on 2026-11-05 and can be checked through the entries above.

This page does not show numbers that cannot be verified on the spot: there are no invented customer cases and no guesswork growth curves.

AI delivery, built asa pipeline you can verify.