NeuralSharks
Start a Project

INSIGHTS · NEURALSHARKS ENGINEERING

Engineering Insights

Field notes on building AI systems, software, and the infrastructure under them: the decisions, trade-offs, working code, and failure modes, with sources.

Browse articles RSS feed

Latest articles

We publish when we have something specific to say about how software gets built and operated. Each article carries its author, dates, and references, so you can judge it on its sources.

FILTER BY DISCIPLINE

Showing 7 of 7 articles

  1. AI Engineering

    The distance between an AI demo and production.

    Why evaluation sets, untrusted-input boundaries, request tracing, and fallbacks belong in the first architecture discussion for an AI product, not the post-launch backlog.

    KEY TAKEAWAYS

    • A demo shows that an idea can work; production requires knowing when it fails, who notices, and what happens next.
    • Build a labeled evaluation set from real inputs before tuning prompts, and track pass rates per category in CI.
    • Treat retrieved documents and tool output as untrusted: scope tools to the user and require approval in proportion to an action's impact.
    NeuralSharks Engineering8 min read1 code example · 1 diagram · 6 references

  2. Architecture

    Draw boundaries around change, not around nouns.

    How to find module boundaries that keep a codebase changeable, enforce them in CI, record why they exist, and recognize when a module has earned its own service.

    IN THIS ARTICLE

    1. 01The question that produces the wrong boundaries
    2. 02Finding boundaries in an existing codebase
    3. 03Start with modules inside one deployable
    NeuralSharks Engineering6 min read3 code examples · 1 diagram · 7 references

  3. Automation

    Every automation needs an exit route.

    Designing automated workflows for timeouts, duplicates, and bad data: failure classification, retries with jitter, idempotency keys, explicit failure states, and operator controls.

    IN THIS ARTICLE

    1. 01The happy path is the easy part
    2. 02Classify every failure before handling it
    3. 03Make every step safe to run twice
    NeuralSharks Engineering6 min read2 code examples · 1 diagram · 7 references

  4. Software Development

    Schema changes without downtime: expand, migrate, contract.

    A step-by-step method for changing a production PostgreSQL schema while the application keeps serving traffic, with the lock behavior that makes each step safe.

    IN THIS ARTICLE

    1. 01Why a one-line migration can take a site down
    2. 02First, make every migration fail fast
    3. 03The pattern: expand, migrate, contract
    NeuralSharks Engineering6 min read5 code examples · 1 diagram · 6 references

  5. Quality Engineering

    Flaky tests are a signal, not noise.

    How to detect flaky tests from CI data, quarantine them without losing coverage, and fix the usual causes: fixed waits, shared state, real clocks, and live network calls.

    IN THIS ARTICLE

    1. 01What a flaky test actually costs
    2. 02Detect flakiness from data, not anecdotes
    3. 03Quarantine with an owner and an expiry date
    NeuralSharks Engineering5 min read5 code examples · 1 diagram · 6 references

  6. Cloud & DevOps

    SLOs that change decisions, not dashboards.

    How to choose service level indicators from user journeys, read a target as an error budget, alert on burn rate instead of thresholds, and agree what happens when the budget runs out.

    IN THIS ARTICLE

    1. 01A target nobody acts on is decoration
    2. 02Choose indicators from user journeys
    3. 03Set the target, then read it as a budget
    NeuralSharks Engineering6 min read2 code examples · 5 references

  7. Product Engineering

    Build for the next release. And the one after.

    How product engineering teams make each release answer a question: release briefs with decision rules, thin vertical slices, flags that separate deploying from releasing, and stable rollouts.

    IN THIS ARTICLE

    1. 01The first release is an instrument
    2. 02Write down the question before the code
    3. 03Slice vertically, and thinly
    NeuralSharks Engineering5 min read2 code examples · 1 diagram · 4 references

START A PROJECT

Facing one of these problems?

Tell us about the system and the constraint you are working within. We will tell you plainly whether we can help, and how we would approach it.

Start a ProjectExplore services