Client stories
These notes come from sittings we were allowed to describe. They mention the work, not a star rating. One of them still wishes we had been firmer. That belongs here too.
Follow-up conversations after a memo
Compute cost review
Farah sat with three months of GPU-hour lines and found a replica we had left up after a pilot in Johor. The memo gave finance a date and an instance name, which ended a month of vague arguing. I still wish the kickoff had spelled out how long our security team would take to export the usage file — that wait ate a week we had not planned.
Ahmad Rahman, Product lead, document answering feature · A Kuala Lumpur fintech
Usage window briefing
We asked for a reading of one noisy November. Priya’s note separated a campaign weekend from a model change the engineering lead had not mentioned to us. The note is short enough that I could walk it in the Tuesday meeting without translating jargon into still more jargon.
Siti Nurhaliza Binti Kamal, Finance partner · Retail group, Bukit Bintang
Cost allocation mapping
Lim would not let us pretend untagged calls were “probably search.” The table uses tagged volume where we have it and a written remainder method for the rest. Search still grumbles. At least they grumble about a page they signed, not about a number I invented in a spreadsheet.
Daniel Tan, Engineering manager · Two internal products on one endpoint
Recurring spend briefing
The monthly note is dull in the best way. When a period is quiet, Priya says so in a paragraph instead of padding. When a period is not, the call is on the calendar already, which is the part we used to skip until the invoice felt embarrassing.
Mei Ling Chong, Operations · Transcription service used by clinics
A fintech team in Kuala Lumpur asked for a compute cost review of a document-answering feature. The invoices showed GPU-hour totals that had not come down after a two-week pilot with a partner in Johor. The usage extract named an instance that nobody in the KL team had logged into since the pilot closed.
The findings memo put the instance name, the last day it should have run, and the monthly cost of leaving it up. Finance used that page in the next budget meeting. The engineering lead used it to write a close-down step into the next pilot runbook.
Two internal products and a batch summariser shared a model endpoint. Call counts favoured the chatbot. Token volume favoured the summariser. Tags existed for two callers and not the third.
The allocation table used tagged volume where it existed and a remainder method written in a paragraph everyone signed. Search still dislikes their share. They dislike a method they saw, which is a different argument from the one they arrived with.