Turning hundreds of dealer group chats into one searchable view of the luxury watch market.

About Luxehouze

Luxehouze trades luxury watches across Hong Kong, Singapore and Indonesia, running as Luxehouze SG Pte. Ltd. and PT. Sinar Wilshire with a team of around 110 people.

In this market the price of any given reference is not published anywhere. It is quoted, minute by minute, inside the private dealer and broker groups the trade runs on. Luxehouze sits inside hundreds of them.

The challenge

What Luxehouze needed to achieve

The trading desk needed to know what a reference was actually going for, right now, without a person having to remember it. Quotes arrive as ordinary prose — a model, a reference number, a condition, a price, a currency — scattered across hundreds of conversations at a combined 20,000 to 100,000 messages a day.

At that volume the constraint stops being access and becomes recall. Everything needed to price and source well is already arriving; nobody can hold it. Buying decisions end up resting on what a trader happens to remember reading.

The challenge

Turning that stream into something searchable is harder than it sounds, because almost nothing about it is uniform:

• The same watch is written a dozen ways. Reference numbers carry colour and condition attributes inside the reference itself, so two strings that look unrelated describe the same item.
• Prices arrive in US dollars, Hong Kong dollars, Singapore dollars and rupiah, often within the same conversation, and have to be comparable.
• The stream is noisy. Dealers repost the same listing across groups and through the day, so raw volume badly overstates real supply.
• Timestamps arrive raw, from three markets in different timezones, which makes “current” ambiguous unless it is normalised first.

The solution

DELOS built the front half of the pipeline: extracting the message stream, converting each post into structured line items, and providing search across them. Luxehouze’s own engineers took it from there, owning the ETL, the data cleansing and the integration into the back office they already run.

Noise removed before parsing. A duplicate-removal engine and template-based pre-filtering strip reposts and spam up front, so parsing spends its budget on messages that carry new information. Filtering before the expensive step is what keeps the economics workable at this volume.

Prose turned into line items. Each post is decomposed into individual listings with the model, reference, condition and price pulled out as fields — including references that carry colour attributes inline. A single message frequently contains several distinct items, and each becomes separately searchable rather than being indexed as one blob of text.

One comparable number. Currency is normalised across USD, HKD, SGD and IDR with conversion applied, and timestamps are reconciled across the Hong Kong, Singapore and Indonesia timezones. Only after both does a cross-market comparison mean anything.

Search over the result. The commercial team queries structured listings instead of scrolling conversations, with new messages folded in on a one-minute cycle so the view reflects the market as it moves.

The system was delivered as containerised services deployed into Luxehouze’s own environment, so the team runs it on their infrastructure rather than depending on ours.