Engineering notes
Written by the people who designed and built the systems, not by marketing. Each note ends with the part that generalises, stated independently of the system it came from. Clients are described, not named, unless they have agreed to be.
Made to choose a code the chart could not support, the model in the medical coding engine we built picked the most common one. So the engine flags the missing detail instead of guessing, and a coder asks the clinician. It has coded more than 19,000 anaesthesia charts.
In the finance review queue we built, the source system, rules and past reviewer decisions answer each transaction first; an LLM handles only what they cannot. It sees only the fields it needs, returns only a client and a reason, and leaves the line for a person when it cannot answer.
For a multi-tenant SaaS platform on a document store, we weighed an account per tenant, a table per tenant and shared tables, and chose shared tables. Sharding tenants across databases and enforcing tenant scope in the data layer bought the isolation back.
We moved a customer success platform from AngularJS to React one dashboard widget at a time, making each widget an independent micro front-end. There was no rewrite and no feature freeze, and the migration finished: the platform is now entirely React.
We run a random cut forest over telemetry from 3,000 EV chargers to catch the faults no rule describes, by flagging units that behave unlike their peers. Each finding is paired with the record of how similar cases were fixed before.
A new digital lender had no repayment history to train a credit model on, so we trained its approval models on about 8,000 decisions its credit officers had made by hand. The model inherits their biases too, and their own decisions are the control group for finding them.
The lending platform we built decides 7,000 applications a day, many from first-time borrowers with no credit history. Four separate models — credit bureau, bank statements, device, SMS — feed an arbiter that knows which evidence is missing and refers a case to a person rather than guessing.
Neither the UPI scheme nor the sponsor bank's core banking system can be tested against on demand. So we built simulators of both, maintained like product code, that answer late, twice or never on command; the awkward paths had run hundreds of times before certification.
Methods in the UPI and IMPS switch we run declare idempotency with an annotation, and one aspect enforces it, so coverage is a grep away. The trap is the obvious check-then-act enforcement, which lets duplicate requests through; a uniqueness constraint claimed first does not.
UPI acknowledges a payment request on one connection and reports the outcome later by calling our switch. How that shaped the switch we run for an ATM operator, and why a missing response calls for a status check, not a retry that could debit someone twice.
Sent one by one, alarms from 3,000 EV chargers teach operators to filter them out, smoke alarms included. The platform we build sends each alarm once, holds self-clearing ones such as dropped connections for fifteen minutes, and sweeps every minute for conditions no event reports.
The broker incentive platform we built processes hundreds of thousands of trades a day in parallel. Order changes a rebate only when a partner crosses a tier, so we reconcile the ten minutes either side of each tier change rather than serialising the whole stream.