The toolchain,
and what it
ruled out.

Chosen Every dependency that matters in machine learning and scientific computing ships wheels for it. It is where the ecosystem is, and arguing with that costs more than it saves.

Instead of Rewriting numerical stacks in a faster language is how you introduce bugs, not how you remove dependencies.

Chosen Where a dropped frame is a defect. The terminal shell must stay responsive with a live feed attached, and a garbage-collected runtime cannot promise that.

Instead of Electron: proven at version one, and proven to have a ceiling once a chart surface, a terminal and a 3D view share a frame budget.

Chosen Rolling volatility and channel extremes, bound in-process. The boundary is drawn at measured cost — everything else stays readable.

Instead of Rewriting whole strategies in C++ would make them faster and unmaintainable, and the strategies are not the bottleneck.

Chosen Decades of statistical libraries with peer-reviewed implementations. It runs as a service so its process lifetime is nobody else’s problem.

Instead of Reimplementing time-series econometrics in Python to avoid a second runtime trades a dependency for a correctness risk.

Chosen Plain JavaScript with no transpilation where the surface is small. No bundler means no lockfile drift and nothing between the source and the browser.

Instead of TypeScript is used where a data model is shared across a team; it is overhead on a single-operator interface with no public API.

Chosen Each is the native language of its layer. Querying through an ORM abstraction, or drawing through a wrapper, costs the thing you went there for.

Instead of Abstractions over SQL make the easy queries shorter and the hard ones impossible.

Chosen Lightning supplies loops, checkpointing and mixed precision for free at these model sizes, which is most of the boilerplate and none of the interesting part.

Instead of JAX was rejected on ecosystem friction with the geometry stack; TensorFlow on research reuse.

Chosen Unlabelled sensor data is abundant and labelled data is measured in minutes. Pretraining first is the only route that respects that ratio.

Instead of Supervised training from scratch needs orders of magnitude more labelled data than any realistic study produces.

Chosen A compact instruction-tuned model with a low-rank adapter fits in consumer VRAM and trains in hours, not weeks.

Instead of Full fine-tuning of a large model buys fluency that was never the constraint, at a cost that was.

Chosen Overnight batch work has no latency requirement, so CPU-capable inference removes a GPU from the production path entirely.

Instead of A GPU serving stack is the right answer for interactive inference and the wrong one for a nightly queue.

Chosen Content is retrieved and ranked from a governed corpus before any prose exists. The corpus is the authority; the model is the writer.

Instead of Generation-first with citations attached afterwards produces fluent text and unverifiable claims.

Chosen Heterogeneous constraints — odometry, heading, barometric floor, dock priors, loop closures — combine in one optimisation with principled uncertainty out the other side.

Instead of A filter processes measurements once and cannot revisit them; a smoother uses the whole shift, which is exactly what a batch pipeline can afford.

Chosen The network predicts displacement and its own uncertainty, so the optimiser knows which segments to trust.

Instead of A fixed noise model treats a confident stride and an ambiguous one identically, which corrupts the whole graph.

Chosen Maps render from a validated schema, so the geometry in the artifact is the geometry that was computed.

Instead of Raster or model-drawn output cannot be checked against its source, which makes it undisplayable under the evidence rule.

Chosen The identical strategy and risk-manager instances run in the live loop and in the backtester. There is no second implementation to drift.

Instead of A separate research implementation is faster to write and guarantees a divergence nobody notices until it costs money.

Chosen The engine cannot tell whether it is reading a live feed or a historical one, so anything a strategy cannot get from the event stream is something the backtest could not have known either.

Instead of Vectorised backtesting is fast and quietly permits lookahead that no live system could reproduce.

Chosen Position size is a function of a multi-factor conviction score under drawdown breakers and cooldowns.

Instead of Fixed fractional sizing treats a marginal setup and a strong one identically.

Chosen The features that produce no return. An engine that dies must not leave a position open behind it.

Instead of Nothing. This layer has no acceptable alternative.

Chosen Transactional, well understood, and rich enough that the interesting logic can live in queries rather than in application loops.

Instead of Document stores were rejected wherever the data is genuinely relational, which is nearly everywhere here.

Chosen Row-level locking with skip-locked is transactional, observable in the same database as everything else, and one fewer service to operate.

Instead of A dedicated queue is correct at thousands of jobs an hour and unnecessary overhead at a hundred a night.

Chosen Chosen specifically for idle economics, accepting weaker native geospatial handling and filling the gap with a dedicated mapping service.

Instead of A richer geospatial engine that cannot pause breaks the cost model that justified the architecture.

Chosen Weights and the data that produced them are versioned together on zero-egress storage, so any published figure traces to an exact input.

Instead of Large-file versioning without dataset semantics stores the bytes and loses the lineage.

Chosen Content that changes when the author changes it does not need a server. Pages are generated from a data corpus at build time and served as files.

Instead of A CMS and a runtime add operational surface for content with one editor.

Chosen Minute-scale analysis cannot live behind a request. Jobs enqueue, a worker drains, and transient failures re-queue rather than failing the user.

Instead of Synchronous request-response turns a rate limit into a lost job and a timeout into a support ticket.

Chosen Where a live chart, a terminal and a 3D view share one frame budget, the shell has to render through the platform graphics API.

Instead of A wrapped web runtime was measured against this requirement and failed it.

Chosen A new capability becomes a server rather than a fork. Slower for the first integration and dramatically cheaper by the tenth.

Instead of Bespoke integrations multiply, and each one is a place for behaviour to diverge.

Chosen Validation at the edges of every pipeline stage, so a source that changes shape fails loudly at ingestion rather than silently at the model.

Instead of Trusting upstream schemas is how a pipeline learns the sensor instead of the task.

Chosen Logs that can be queried rather than read. A long pipeline will fail partway through, and the question is always which record and which stage.

Instead of Printed strings are unusable at the point you actually need them.

Chosen Invariants over generated inputs catch the edge cases that hand-written examples do not, which is most of them in DSP and closure code.

Instead of Example-based tests confirm the cases the author already thought of.

Chosen Every stage re-runnable without corrupting what came before, because rate limits, layout changes and exhausted quotas are certainties.

Instead of Designing for a clean single pass produces a pipeline that is nursed rather than operated.