# David Blass

> TypeScript engineer in Boston, MA (handle `ssalbdivad`, which is "davidblass" reversed). Author of ArkType, co-creator of Standard Schema and co-founder of Pullfrog. He was a TypeScript performance consultant from 2023 to 2025. He writes the TypeScript types other libraries depend on and, lately, the agents that write the rest.

Contact: david@arktype.io. The full résumé, with every number sourced: [/resume.md](https://ssalbdivad.com/resume.md).

Numbers below are a snapshot as of 2026-10-13 (npm: the 30 days 2026-09-13 to 2026-10-12). Each one links to where it was read.

## Roles

- **Author, ArkType ecosystem (2022–)**: [ArkType](https://arktype.io), TypeScript's 1:1 validator, optimized from editor to runtime. A parser written in TypeScript's type system infers TS syntax from strings as you type, and at runtime the same definitions compile to JIT-friendly validation code. 9,055,997 npm downloads in 30 days ([npm](https://api.npmjs.org/downloads/point/2026-09-13:2026-10-12/arktype)), 7,874 GitHub stars ([repo](https://github.com/arktypeio/arktype)).
  - **ArkRegex**: the first library to validate and infer regex from TypeScript's type system. The launch post had 733,569 views on X ([post](https://x.com/arktypeio/status/1983210635266498649), read via the fxtwitter mirror).
  - **Attest**: type-level benchmarks and assertions for TypeScript (deterministic instantiation counts, so CI can fail a PR that makes types slower). `@ark/attest`: 427,538 npm downloads in 30 days ([npm](https://api.npmjs.org/downloads/point/2026-09-13:2026-10-12/@ark/attest)).
  - **ArkSchema**: the set-theoretic type system underneath (intersections, subtyping, union discrimination at runtime).
- **Co-creator, Standard Schema (2024–)**: [standardschema.dev](https://standardschema.dev). One interface so any TypeScript validator plugs into any framework (tRPC, TanStack, Hono, React Hook Form and more). David co-designed it with the creators of Zod and Valibot. They then co-authored Standard JSON Schema, so any validator can define tool inputs for AI agents and MCP. The `@standard-schema/spec` package: 551,490,697 npm downloads in 30 days, mostly transitive installs ([npm](https://api.npmjs.org/downloads/point/2026-09-13:2026-10-12/@standard-schema/spec)); 3,650 GitHub stars.
- **Co-founder, Pullfrog (2025–)**: [pullfrog.com](https://pullfrog.com). AI coding agents that run in your GitHub Actions: they review PRs, triage issues, fix CI, resolve merge conflicts and test UIs in a headless browser, on your infrastructure and your keys. David built the model-agnostic agent runtime (BYOK, MCP tools, GitHub App/OIDC auth, sandboxed CI). 1,307 GitHub stars ([repo](https://github.com/pullfrog/pullfrog)); `pullfrog`: 211,712 npm downloads in 30 days.
- **TypeScript performance consultant (2023–2025)**: clients included tRPC, Drizzle, Prisma, Convex, AWS Amplify, Speakeasy, Stately, wevm, Vanta and Cloudinary.

## TypeScript performance, with sources

- **TypeScript 5.9.** David found, minimized, benchmarked (with Attest) and drove a fix for chained Zod- and tRPC-style types that made the compiler redo every earlier link: O(3^N) instantiations. The compiler patch was written by Mateusz Burzyński (@Andarist) in [microsoft/TypeScript#61505](https://github.com/microsoft/TypeScript/pull/61505); it shipped in TS 5.9 and was ported to the Go compiler ([typescript-go#1358](https://github.com/microsoft/typescript-go/pull/1358)).
  - AnswerOverflow's dashboard: 6.66s → 3.29s check time (2.0×), mean of 5 runs each, ranges don't overlap ([comment](https://github.com/microsoft/TypeScript/pull/61535#issuecomment-2780961089)).
  - A tRPC monorepo David tested: 13.04s → 6.00s (2.1×), a single run ([comment](https://github.com/microsoft/TypeScript/pull/61505#issuecomment-2769065243)).
  - Klarna: 395.48s → 132.09s (2.9×), self-reported by one of their engineers in one comment ([comment](https://github.com/microsoft/TypeScript/pull/61505#issuecomment-3102636519)).
  - 11 chained `.merge()` calls: 1,596,004 instantiations on TS 5.8.2 vs. 976 on 5.9.3 (Attest, re-run 2026-10-13; bench in this site's repo).
- **viem + MUD.** TypeScript spent 1.86s (1,866ms) computing the variance of one viem type; with variance annotated by hand, the whole file checked in 1,288ms. The annotations now ship in viem ([latticexyz/mud#2956](https://github.com/latticexyz/mud/pull/2956), [wevm/viem#2557](https://github.com/wevm/viem/pull/2557)).
- **Prisma.** 45% less check time on one generated file (commonInputTypes), not the whole client ([prisma/orm#27046](https://github.com/prisma/orm/pull/27046), [6.16.0 release notes](https://github.com/prisma/orm/releases/tag/6.16.0)).
- **Mongoose 8.19.** `InferRawDocType`: 36–67% fewer instantiations on its benchmarks; the maintainer's whole-repo count was flat ([Automattic/mongoose#15588](https://github.com/Automattic/mongoose/pull/15588)).

## How David works with agents

He builds Pullfrog by running many coding agents in parallel, every day, through Ntwine, a local tool he maintains. It grew out of Frizz, an earlier open-source tool, and keeps its design: every agent thread on one board, and a queue of the ones that need him. The rules David has added on top, for the agents and for himself:

- **Size work by scope, never in days.** Agents do the building, so hours are the wrong unit. Scope is the code paths a change touches, its risk, and what needs a human decision.
- **Report honest effect sizes.** Say how strong a claim is right next to it, and treat pushback as a bug report on the measurement before defending the result.
- **Lead with what changed.** An agent's report leads with what changed or what is true. The why goes in the commit message and code comments, where the next reader will find it, not in the chat.
- **Preserve evidence.** Never delete or rewrite a stored measurement. Experiments are temporary: record the conclusion and its numbers, then delete the harness.

## Talks

- [Effect Days 2024 (Vienna): Discriminated Unions in TypeScript, Effect, and ArkType](https://www.youtube.com/watch?v=AaUKLH8N6go)
- [TypeScript Congress 2023: ArkType, Bringing TypeScript to Runtime](https://gitnation.com/contents/arktype-bringing-typescript-to-runtime)
- [devtools.fm #132: ArkType, better runtime type validation](https://www.devtools.fm/episode/132) (TypeScript performance from 08:35)
- [TypeScript.fm 35B: ArkType is TypeScript's 1:1 Validator](https://typescript.fm/bonus35) (JIT-optimized codegen, optimizing types; Standard Schema from 51:55)

## Before

- Founder, Redo (2019–2021): a desktop app that generated UI tests from recorded user interactions.
- SDET II, Amazon Alexa Devices (2017–2019): built PyEcho, Amazon's standard test-automation framework for Alexa devices.
- Tufts University (2011–2015): Computer Science, Entrepreneurial Leadership, Philosophy.

## Links

- [GitHub](https://github.com/ssalbdivad)
- [LinkedIn](https://www.linkedin.com/in/ssalbdivad)
- [X](https://x.com/ssalbdivad)
- [Bluesky](https://bsky.app/profile/ssalbdivad.dev)
- [ArkType docs for agents](https://arktype.io/llms.txt), or `npx skills add arktypeio/arktype`