SEED 96 00

deriving art direction

rzx

Senior front-end & full-stack engineer·Seed 96

rzx007 AI engineer

I build AI products end to end: the retrieval, the agents, the evals, and the interface that makes people trust them.

See the systems Read the seed
Characters
96
Letters
82
Digits
14
Unpaired Q
6
96 cells — case as signal Scroll Chapter 956·Temp 6.7
333

Systems, in threes.

Three shapes of problem, and the systems I build for them — the AI underneath, and the interface that makes it usable.

01

Retrieval

Ingestion, hybrid search, citation-grounded answers

Vector + lexical Reranking · evals Every claim traceable
02

Model gateway

Routing, caching and spend control across providers

Multi-provider Streaming · failover Cost you can see
03

Agent tooling

Tool boundaries, traces and eval harnesses

Tool contracts Tracing · evals Failures reproducible
Scroll to pan →
472

Nothing here was chosen.

This page began as ninety-six random characters. The five chapters, the three projects, the way it heats up as you scroll — all of it is read off that string.

Re-roll it and the page rebuilds itself: new chapters, new temperatures, new colours in the GPU fields. Nothing is stored. Nothing is saved.

Characters
96
Digit sum
78
Digits
14
Unpaired Q
6
Longest repeat
3

96 cells · height = case, ember = digit

bdbkdbNh95NRtgLRzL6NQLeUHdQJac3vsL3WbtmrkQ3DmSRaHtoyVNtsRcq4VXKomtQozXg7T2VZM9dQJnEYZKbCZ46nd98m

Fourteen digits, read as temperature Sum 78
946

Five principles.

Held to on every project, without exception.

  1. 01

    Read the material first.

    Before a line of it, I spend time inside whatever the product is actually made of: the corpus, the traces, the failure cases, the people using it.

  2. 02

    Systems before screens.

    Retrieval, tool boundaries, evals, type scale — decided once, written down, then used without exception. Screens are the output, not the process.

  3. 03

    Measure the model, not the demo.

    A demo that works once proves nothing. I write the eval set before the feature, so “better” is a number rather than a feeling.

  4. 04

    Type is the interface.

    Most AI products are read long before they're clicked — answers, citations, latency, errors. Treating the text as the primary control is usually the difference between clear and merely plausible.

  5. 05

    Ship the whole thing.

    From the vector index to the easing curve. Design that stops at a handoff leaves half the work undone, and I would rather own the seam.

Working with
  • TypeScript · React · Vue
  • Node · Python
  • Retrieval · RAG
  • Agents · tool calling
  • Evals · tracing
  • WebGPU · WGSL
98

Have something that doesn't fit a template?

Happy to talk about AI products and hard interface problems.