shipwithjev

The shipwithjev press · 175 posts · page 2 of 5

Blog.

What the builds in the catalog add up to. Every post goes through the same gate: every number is as reported by its author.

  • Questions 31
  • Builds & people 30
  • Comparisons 24
  • Recipes 23
  • Evals & judging 12
  • Triage & routing 11
  • Classification 11
  • Guardrails 10
  • Cost 7
  • Calibration 7
  • Monitoring 6
  • Essays 3
Fig. 00 · The pressIn, judged, stamped, out
  1. FIG. 136Cost

    Builds That Cost Under a Dollar

    Cheap AI projects with receipts: Jev builds that ran for under a dollar, from a hundredth of a cent to 67 cents. Builder-reported, all linked.

    Sep 30, 2026Read →

  2. FIG. 135Builds & people

    The Weirdest Jev Builds (Complimentary)

    Weird AI projects built on Jev: semantic SQL, a traffic light that does nothing, paintings from probabilities. Unusual builds, receipts linked.

    Sep 30, 2026Read →

  3. FIG. 134Triage & routing

    The Best Triage Builds in the Directory

    Jev triage examples worth copying: inbox sorters, escalation lanes, log and ticket triage, each with the builder's reported numbers linked.

    Sep 30, 2026Read →

  4. FIG. 133Builds & people

    The Best Jev Game Builds So Far

    A curated Jev games list: Doom, Slay the Spire 2, Wordshift, Tetris and the honest losses, with each builder's reported numbers linked.

    Sep 30, 2026Read →

  5. FIG. 132Monitoring

    Where to Check Jev Status and Uptime

    Is Jev down? Where to check Jev status and uptime, how to tell an outage from your own bug, and the fallback policy every pipeline should have.

    Sep 30, 2026Read →

  6. FIG. 131Questions

    Can Jev Score Essays?

    Can Jev do AI essay scoring? It can check rubric criteria fast, but essays stay human-graded. The short-form vs essay line and the formative rule.

    Sep 30, 2026Read →

  7. FIG. 130Questions

    Is Jev HIPAA Compliant?

    Is Jev HIPAA compliant? We found no public claim either way. The BAA question to ask, how to design around PHI, and the audit posture to keep.

    Sep 30, 2026Read →

  8. FIG. 129Questions

    Can Jev Handle PII?

    Can Jev handle PII? It can process it, but the better question is how little you send. The minimization pattern, a PII-finding build, and what to ask.

    Sep 30, 2026Read →

  9. FIG. 128Questions

    Can I Fine-Tune Jev?

    Can you fine-tune Jev? No documented fine-tuning path exists. Your questions are the tuning surface: boundary clauses, optimizers, and layers on top.

    Sep 30, 2026Read →

  10. FIG. 127Questions

    Is Jev a Company or a Model?

    Is Jev a company? No: Jev is a decision model; TypeSafe AI is the company that makes it. Who owns what, where to get access, and the name clash.

    Sep 30, 2026Read →

  11. FIG. 126Questions

    Who Uses Jev? The Builder Census

    Who uses Jev? A census of cataloged builds by category and source, from solo GitHub tinkerers to Metaview, Vercel, and LangChain integrations.

    Sep 30, 2026Read →

  12. FIG. 125Comparisons

    Jev vs Jev Engineering: The Term Confusion

    What does jev engineering mean? Jev is TypeSafe AI's decision model; jev engineering is the craft of writing questions it can rule on.

    Sep 30, 2026Read →

  13. FIG. 124Questions

    Does Jev Learn From My Feedback?

    Does Jev learn from your corrections? No memory between calls. Where feedback actually goes: better questions, logged verdicts, and a model you train.

    Sep 30, 2026Read →

  14. FIG. 123Questions

    Can Jev Translate? No, and What To Do Instead

    Can Jev translate text? No: the decision model picks answers, it doesn't write them. What to pair it with, and why you rarely need to translate first.

    Sep 30, 2026Read →

  15. FIG. 122Questions

    Can Jev Summarize Text? (No; Here's the Split)

    Can Jev summarize? No: Jev does no text generation. The split that works: a chat model writes the summary, Jev makes the decisions around it.

    Sep 30, 2026Read →

  16. FIG. 121Questions

    Is Jev Deterministic?

    Is Jev deterministic? No: the same input can draw a different verdict. Why, how much it matters, and the cache-and-log rule that fixes replay.

    Sep 30, 2026Read →

  17. FIG. 120Recipes

    Cost Monitoring for Verdict Pipelines

    LLM cost monitoring for verdict pipelines: what to meter per call, the ratio that catches runaway loops, and API spend alerts worth setting.

    Sep 30, 2026Read →

  18. FIG. 119Recipes

    Human-in-the-Loop That Humans Don't Hate

    Human in the loop design for verdict pipelines: evidence-attached queues, sane volume, and review queue UX that keeps reviewers sharp instead of numb.

    Sep 30, 2026Read →

  19. FIG. 118Recipes

    Building Gold Sets That Stay Gold

    How to build a gold standard dataset for LLM judges: sourcing, blind labeling, adjudication, and the rules that keep eval ground truth from rotting.

    Sep 30, 2026Read →

  20. FIG. 117Monitoring

    Monitoring Judge Drift

    Model drift monitoring for verdict pipelines: track verdict distribution shift, escalation rate and a frozen canary set to tell which drift you have.

    Sep 30, 2026Read →

  21. FIG. 116Guardrails

    Instruction Bleed and How to Beat It

    Instruction bleed makes LLM judges rule on what text mentions, not what it does. How to spot it, test for it, and write questions that resist it.

    Sep 30, 2026Read →

  22. FIG. 115Calibration

    Self-Preference Bias: The Deep Dive

    LLM self-preference bias, in depth: why judges favor their own family's outputs, where it hides in evals, and how cross-family judging fixes it.

    Sep 30, 2026Read →

  23. FIG. 114Calibration

    Per-Choice Probabilities, Explained

    LLM probability output, explained: how to read per-choice scores, margins and the unclear band, and when calibrated probabilities are worth trusting.

    Sep 30, 2026Read →

  24. FIG. 113Recipes

    Choice Set Design: Labels That Judge Well

    Classification label design for decision models: build answer sets that are exclusive, exhaustive, and action-shaped, so verdicts stop wobbling.

    Sep 30, 2026Read →

  25. FIG. 112Comparisons

    Batch vs Real-Time Verdicts

    Batch LLM processing or real-time verdicts? Pick by when the answer must exist, how cost scales with clock or items, and what builders report.

    Sep 30, 2026Read →

  26. FIG. 111Triage & routing

    Escalation Design Patterns

    Eight LLM escalation patterns for decision pipelines, from confidence gates to never-approve gates, with builder receipts and the anti-patterns.

    Sep 30, 2026Read →

  27. FIG. 110Recipes

    Question Versioning in Practice

    Prompt versioning for judge questions: what counts as a breaking change, how to tag verdicts, and how to roll out a reworded question safely.

    Sep 30, 2026Read →

  28. FIG. 109Monitoring

    Verdict Logging and Audit Trails

    Build an AI decision audit trail: what each verdict log record needs, what to leave out, and how to answer who decided and why. Not legal advice.

    Sep 30, 2026Read →

  29. FIG. 108Calibration

    Calibrating a Judge: The 100-Case Method

    LLM calibration in one afternoon: the 100-case method for verdict agreement testing, reading disagreements, and checking probability buckets.

    Sep 30, 2026Read →

  30. FIG. 107Calibration

    Confidence Thresholds: The Complete Guide

    Confidence threshold LLM guide: how to set probability cutoffs from labeled data and error costs, with two-sided bands and builder receipts.

    Sep 30, 2026Read →

  31. FIG. 106Comparisons

    Decision Model vs LLM: The Category Explainer

    Decision model vs LLM: what a decision model is, how it differs from a chat LLM, and the builder receipts that show where each one belongs.

    Sep 30, 2026Read →

  32. FIG. 105Comparisons

    Decision Models vs Agent Frameworks: Different Layers, Same Stack

    Agent framework vs model is a layer mix-up. Where a decision model like Jev plugs into LangChain, Pydantic AI and Composio, with receipts.

    Sep 30, 2026Read →

  33. FIG. 104Comparisons

    Jev vs Human Review: The Real Math

    AI vs human review cost, done honestly: the formula, builder-reported verdict prices, and why the win is coverage, not headcount.

    Sep 30, 2026Read →

  34. FIG. 103Comparisons

    Jev vs a Fine-Tuned BERT: Which Classifier Should You Ship?

    BERT vs LLM classification, specifics only: labels, 512 tokens, retraining, and when a fine-tuned small model should inherit a Jev pipeline.

    Sep 30, 2026Read →

  35. FIG. 102Comparisons

    Jev vs Rules Engines

    Rules engine vs AI: when rules beat models, where Jev's verdicts take over, and the rules-floor pattern that keeps both honest in one pipeline.

    Sep 30, 2026Read →

  36. FIG. 101Comparisons

    Jev vs Embeddings: Ruling vs Resembling

    Embeddings vs LLM classification: embeddings measure likeness, Jev rules on a question. When each wins, and why the best stacks use both.

    Sep 30, 2026Read →

  37. FIG. 100Comparisons

    Jev vs Llama: Hosted Verdicts vs Self-Host

    Jev vs Llama for open model classification: what hosted verdicts cost, what self-hosting costs, and when running Llama yourself wins.

    Sep 30, 2026Read →

  38. FIG. 99Comparisons

    Jev vs Kimi K3: Partners, Not Rivals

    Jev vs Kimi: a decision model and a general model do different jobs. In the Kimi K3 cascade fraud build, the pair scored 96/100 for ~$0.07.

    Sep 30, 2026Read →

  39. FIG. 98Builds & people

    Jev for Nonprofits: Triage on a Zero Budget

    AI for nonprofits on a tight budget: donor email triage and volunteer routing for cents, as reported, with humans on every gift decision.

    Sep 30, 2026Read →

  40. FIG. 97Builds & people

    Jev for Educators: Formative Only, Loudly

    AI for teachers, done responsibly: classroom feedback automation for practice work and item analysis, with every grade that counts left to you.

    Sep 30, 2026Read →