John Lindquist joins Claire Vo on How I AI to demonstrate eight ways to use TypeSafe AI’s Jev as a fast decision step. The recurring pattern is narrow and useful: turn messy language or state into a choice, score, or yes/no probability, then let ordinary code decide what happens next.

1. Streaming voice control for a to-do list

  • Continuous dictation adds, removes, completes, and reprioritizes tasks without waiting for a submit button.
  • The app chains several narrow judgments: whether the speech is actionable, which task it refers to, and which operation should run.
  • A moving transcript window accumulates enough context to act, emits a structured event, and then starts a fresh window.

2. Plain English to a known function

  • A request such as “What’s in my cart?” maps onto a predefined function such as getCart.
  • The useful constraint comes first: application code defines the valid functions, while Jev maps variable human language onto that bounded set.
  • The same pattern can route API calls, MCP tools, product commands, or agent actions.

3. Record matching before deterministic merging

  • Jev compares messy company or contact records and scores likely duplicates; ordinary merge code handles the records that clear a threshold.
  • Ambiguous matches can go to a human or a generative model for explanation instead of being merged automatically.
  • Lindquist calls the process millisecond-fast for large datasets and discusses scales of 60,000 or 600,000 records, but the visible example is small and no benchmark is supplied.
  • A model’s 99% confidence is not 99% production accuracy until it has been calibrated against representative labeled data.

4. Hierarchical routing through an application

  • A command entered anywhere in a product can select the relevant sub-application, target object, and operation in successive steps.
  • The architecture is a stack of small routers—application → tool → entity → action—not one unconstrained request to infer an entire workflow.
  • For larger search spaces, Lindquist suggests clustering candidates into groups of roughly 30–60, ranking within each group, and caching stable results.
  • Jev is fast, not latency-free; independently scoring a whole corpus can still be too slow for an interactive interface.
  • Jev ranks legal chess moves, explores possible replies, and chooses from the reduced decision tree.
  • Lindquist reports moves under one second, with Jev averaging 10× faster and 4× cheaper than his comparison model.
  • This is a demo, not a controlled benchmark: the opponent is an unidentified free, low-reasoning OpenRouter model, with no repeated trials or move-quality metric.
  • The more transferable pattern is broad, cheap pruning followed by a stronger reasoning model when the remaining decision is genuinely difficult.

6. Interfaces as finite action sets

  • A browser looks like an open-ended canvas, but its DOM exposes a finite set of links, buttons, fields, and controls that can become choices.
  • Games and controllers have the same shape when the legal actions are known in advance.
  • Choosing among known controls fits Jev; judging layout, color, or visual usability still needs a multimodal generative model.

7. Wikipedia routing and multi-agent coordination

  • A Wikipedia pathfinder scores the links available on the current page, picks a next hop toward “Philosophy,” and repeats at every node.
  • A grid simulation assigns three workers separate jobs while Jev picks each next movement and steers them away from shared spaces.
  • Both demos recalculate from current state instead of generating one fixed plan at the beginning.
  • The grid illustrates a coordination pattern; it does not establish production-scale scheduling, deadlock prevention, or safe parallel code editing.

8. A real-time presentation coach

  • Live speech is compared with required talking points so the interface can mark covered items and warn about omissions.
  • The system keeps a rolling speech window, records structured coverage events, and discards processed text.
  • The same state could advance slides or surface a prompt at the right moment.
  • The demo does not measure transcription errors, false positives, false negatives, or performance over long sessions.

The architecture underneath the demos

  • Treat Jev calls as probabilistic if/else or switch statements inside ordinary software, not as autonomous workers.
  • When one broad judgment is unreliable, split it into successive narrow decisions or change the question from yes/no to a choice or score.
  • Send low-confidence cases to a larger model or a human. Keep generative models for explanations, brainstorming, visual interpretation, and undefined action spaces.
  • Typed output prevents malformed or out-of-schema answers; it does not guarantee that the selected answer is correct.

“You put in unstructured sentences and data, but then you get structured data out.” — John Lindquist

What the numbers do—and do not—establish

  • Lindquist says hundreds of demos cost him $0.73. Vo reports about $0.04 for 20,000 records and about $0.40 to process five gigabytes of JSON. The video does not provide billing logs, schemas, token counts, or reproducible workloads.
  • “Completely free through Vercel AI Gateway” is not a permanent pricing guarantee; free credits and model pricing can change.
  • Lindquist had used Jev for only a matter of days when this was recorded. These are useful product sketches, not mature production evaluations or independent tests of TypeSafe’s latency and cost claims.