Client stories
Rows people could argue with
These notes name the work that was done. One of them names a delay we caused. That belongs here as much as the rest.
They sat with our composer logs for a week and handed back a table our product lead could actually read. The nightly rewrite job — which none of us had put on the original list — was the row that changed the conversation.
The first draft assumed we had request IDs on every line. We did not, and that cost two extra days while they rebuilt sequences from timestamps. I would have preferred they refused to start until we added the field. The route-by-route picture, once it arrived, was still the first one our ops lead did not argue with.
We thought we were ready for a full assessment. The instrumentation review said our traces could not tell a retry from a first attempt. That was an uncomfortable afternoon and the right one.
Priya kept the room on the inventory when we started sliding into prompt opinions. The one-page record the next morning matched what people had actually said, including the two points we did not settle.
The snapshot after we killed the save-hook rewrite was short, slightly dry, and exactly what we needed. It did not try to congratulate us.
A longer note: the Subang document assistant
Hafiz’s team had a writing product with a visible composer and a less visible nightly rewrite. The first inventory they handed us listed the composer and a “suggestions” panel. It did not list the job that re-summarised every open record, including records whose text had not changed.
The assessment window was twelve days of traces plus a provider invoice used only as a cross-check. Waiting completions were almost entirely the composer. The nightly job was not waiting on anyone and was the louder row. The walkthrough spent most of its ninety minutes on whether that job was still wanted, not on model choice.
We did not stay to redesign the job. That was not the commission. The useful residue was a table both engineering and product could point at without translating.
A longer note: the help desk without request IDs
Mei Ling’s internal help desk could grant log access but the lines had timestamps and route names, not request identity. We should have stopped at intake. Instead we tried to reconstruct sequences, lost two days, and said so in the report. The later picture of agent-facing replies versus classify-on-open was usable. The inferred rows were marked. We now refuse to start a full assessment when identity is absent, unless the team agrees in writing that a portion of the picture will remain inferred — or books an instrumentation review first.