The AmplifierSoftware · Chapter 9
Arjun's fraud-detection proposal
The second attempt, three weeks later. The same framework, and a different person inside it: Step 1 in fourteen minutes, a structural problem caught in forty-five.
Planned for release phase 2. The page is complete as a draft; interactive parts run in the browser and do not yet save to an account.
In brief
This case is laid out against the eight steps of the framework. Jump to Step 1 to see how the end state was defined, or to Step 8 to see what integration caught.
Define the End State
In a normal weekday text editor, not a separate notebook — the discipline has moved into the work. Four audiences named before he is halfway down the page: Karan carrying it up the chain, the director's office assessing it, the project lead who will have to build it, the client's CTO who needs a vendor that understands his fraud landscape. Fourteen minutes.
Map the Flow
Twenty-two tasks by hand. AI review catches an RBI requirement that thresholds be configurable rather than hard-coded, and a dependency on the client's data-warehouse format he had assumed was standard. A multi-region failover suggestion declined after a back-and-forth: single-country bank, overkill.
Assign Tasks
Just over an hour. Most tasks land cleanly first time; where they do not, he decomposes without resistance, on the assumption that ambiguity is always information.
Select the Tool
Long-context conversational tool for the flow review and drafting; nothing exotic. The selection is quick because it is no longer a decision he has to remember to make.
Prepare and Direct
Criteria written before output, as before — including, for the implementation timeline, "based on the chosen technical approach". That phrase becomes the structural problem.
Execute
Steps 1–3 finished by Tuesday lunch, where the migration had taken until Wednesday evening. Not faster from rushing; faster because the friction of remembering is gone.
Evaluate
Wednesday: the cross-team meeting. Sridhar asks what happens when a batch boundary splits a coordinated transaction sequence. The pause arrives — the same pause that broke Abhi — and then Arjun thinks: sliding window, overlapping batches, deduplicate at output, double the compute, manageable at this volume, currently a refinement on slide twelve and it should be the main approach. "That's the right answer."
Integrate and Stress-Test
Thursday: the technical section now assumes bursty writes; the timeline still assumes steady writes; compute is wrong by thirty percent. Not a small fix. He traces it: a Step 3 criterion pointed at "the chosen approach" as a placeholder, and the timeline was never re-run when the approach changed. Re-brief with the current approach, re-run, re-evaluate. Forty-five minutes from noticing to fixed. On the migration the same class of problem took a day and a half.