Software Engineer

Constraint

Streaming webcam frames to a server for inference puts the whole video payload on the network path.

Decision

MediaPipe can extract hand landmarks in the browser, so what actually crosses the network is a small vector of coordinates rather than an image. Normalising those features also decouples accuracy from lighting, which a server-side pipeline would have had to correct for anyway.

Constraint

At a sub-100ms per-frame budget, HTTP request overhead is a meaningful fraction of the budget.

Decision

Every polled request pays connection setup again, which makes latency bursty rather than predictable. A persistent WebSocket removes that cost entirely and makes the frame budget something you can actually reason about, which matters more for an interactive demo than raw throughput does.

Constraint

The model was accurate enough and too slow to feel responsive.

Decision

Batch processing and quantisation lifted throughput 40% and cut the model to a quarter of its size, at under 2% accuracy loss. For something a person interacts with in real time, the delay is the thing they notice; the last two points of accuracy are not.

Constraint

Several people editing one shared ledger at once, on mobile clients that may be offline mid-edit.

Decision

Mutable balances need locking, and locking assumes connectivity and a coordinator — neither of which a phone reliably has. Storing every expense as an immutable event and folding the log into a balance removes the race condition rather than defending against it. There is nothing to contend over, so there is nothing to lock.

Constraint

Settling a shared ledger pairwise generates far more transfers than the group actually needs.

Decision

Who owes whom is a graph problem, not an accounting one. A minimum cash flow pass runs in O(n log n) and reduces any group of n people to at most n-1 transactions, which is the floor. The naive pairwise version is correct and asks people to make payments that cancel each other out.

Constraint

High-traffic endpoints were slow and a cache was the obvious first move.

Decision

A cache in front of an N+1 query hides the problem and adds an invalidation bug to maintain. Restructuring the indexes and eliminating the N+1 patterns made the underlying reads fast, which is a smaller system to reason about than the same reads plus a cache layer.

Constraint

Four parallel document-ingestion streams were serialising behind synchronous I/O.

Decision

The work was almost entirely waiting — on the embedding API and on the vector store — so threads were sitting idle rather than computing. Async I/O in FastAPI let the streams interleave their waiting, which is where the bottleneck actually was.

Constraint

A document that fails to embed still leaves the pipeline running, just with worse retrieval.

Decision

Silent degradation is the worst failure mode in a retrieval system: nothing errors, results simply get less relevant, and you find out weeks later from a user. Validating at ingestion turns an invisible quality problem into a visible failure someone can fix.

Constraint

A model trained to agree with the triage nurse learns to reproduce the nurse's misses, and in the 2022 NHAMCS data 45.8% of critical outcomes came from patients triaged level 3, 4 or 5.

Decision

Optimising for agreement optimises for the wrong target. The engine is trained on outcomes rather than on assigned levels, and it keeps watching patients while they wait — because triage is a moment and deterioration is a process. Repeat observations re-run the assessment and escalate on their own.

Constraint

Under-triage and over-triage are not equally bad, but taking the most likely level treats them as if they are.

Decision

Missing a critical patient can kill them; escalating a well one costs a bed and some attention. So the level is chosen from the probability distribution with an explicitly asymmetric loss function rather than by argmax. The price of that choice is stated rather than hidden, because a triage assistant that over-triages everything gets worked around within a week.

Constraint

A heart rate of 152 is unremarkable in a one-year-old and an emergency in a seventy-year-old.

Decision

Passing age in as one more column asks the model to learn that relationship from scarce paediatric data. Instead vital signs enter as deviations from ten age bands, estimated from the survey's own low-acuity discharged population so the reference describes well physiology at that age rather than the ED case-mix. Where a band is too thin to estimate the artifact records that it fell back to published values, and the assistant lowers its own confidence there.

Constraint

An attribution model is free to report that traffic contributed negatively to the air quality in a ward, which means nothing to the person deciding where to send an inspector.

Decision

Ordinary least squares will happily produce that, and it fits better for it. Non-negative least squares constrains every coefficient at or above zero, so the breakdown stays physically interpretable and can be defended when it sends someone somewhere. Fit quality is worth less here than being able to explain the answer.

Constraint

A forecast can look impressive in isolation and still be worse than assuming the next hour looks like this one.

Decision

Persistence is the honest baseline for anything time-series, so it is in the test suite: mean RMSE has to come in at least 10% under it or the build fails. The evaluation report leads with the caveat that the holdout is synthetic rather than with the headline improvement, because a number nobody qualified is a number that will be quoted back wrongly.

Constraint

An MCP server speaks JSON-RPC over stdio, so a single stray print statement corrupts the stream and the client disconnects.

Decision

There is no way to have both a debug log and a working server on the same channel. Every diagnostic goes to stderr, which the host shows and the parser never reads. The README is also truncated to 2,000 characters before it is returned — context is the scarce resource for an agent, and spending it on the back half of a readme leaves less for the task.

Constraint

A copied server object means two objects holding the same file descriptor, and the second destructor closes a socket that is already gone.

Decision

The copy constructor and copy assignment are deleted outright and move semantics provided instead, so ownership transfers rather than duplicates. The compiler then refuses the bug at the point it would be written, which is cheaper than finding it as an intermittent failure once several clients are connected.