AEGIS INFORMATION TECHNOLOGY LIMITED · HONG KONG
We build AI systems for construction, finance and operations. Every number traces back to its source, every approval leaves a trail, and nothing is called done until it is proven.
OUR PRINCIPLE
Every step has an owner and a state. Drafts, reviews and approvals tiered by contract value, so no work disappears into a chat thread.
Every claim links to its source: the photo, the clause, the original file, the exact words of the person who decided.
Numbers are computed by code, explained by AI and checked twice before anyone signs.
ONE RECORD, THREE STATES
AI drafts the record from whatever the site sends. A person confirms it. From then on, every downstream number points back to it.
“Poured the east slab on 2F today, about forty cubes, pump truck left at four.”
Sample data for illustration.
PRODUCTS
Site diaries to settlement, for contractors.
Ask the books. Get a number you can trust.
Tender formats you can verify.
One gate for every AI call in the company.
A night watch for Victron energy systems.
LAB · HOW WE BUILD
Show the evidence, or it didn't happen. These are the tools we built to make that true inside our own engineering.
Our AI agents share one ledger. Records are only ever appended. Originals are stored with their hash. A value that changes is superseded, never overwritten, so the old one stays on record. A decision is kept with the exact words of the person who made it, and standing rules are put back in front of the agent on every turn.
An agent cannot report a release as live without a recorded check against the real system. Before any deploy, we ask what production is actually running. Once, production was 34 commits ahead of the main branch. The gate caught it, so we merged first instead of overwriting work that already worked.
One lead model writes the contract. Writer models from different vendors build in parallel. Test writers work from the spec alone and never see the code. Reviewers look for one kind of failure each, refuters must reproduce a finding before it counts, and mutation tests check that the tests actually bite. Every run is logged.
Sixteen photos of a real subject become a generated mesh. Agents rig and animate it in Blender, render frame checks and confirm the result opens cleanly, then export glTF for the web. Each version is kept with its own acceptance notes.
SELECTED WORK
Clients are not named. Each project is described by sector and scope.
Research institute
Logistics hub operator
Development zone operator
Telecom operator
Transport network operator
Water utility group
Systems-integration team
Language school · pilot
HOW WE DELIVER
We sit with the people who do it and find where evidence gets lost.
Your real templates and document formats, so the output is something you would actually sign.
On your premises, in your cloud, or hosted in your region.
Every release is checked against what production actually runs, with a backup and a rollback point first.
[CONTACT EMAIL]
OFFICE
Hong Kong
[REGISTERED ADDRESS]
WE WORK IN
English · 繁體中文 · 日本語 · العربية