ML in the fleet

What machine learning actually runs in the subagent*.com family — a short, verified inventory, not a manifesto. Two production inference surfaces, two offline data-engineering skills, and one live writeup. That's the whole list.

Inference: Cloudflare Workers AI

Exactly one model id appears in the fleet's production code: @cf/zai-org/glm-5.2 (Z.ai's GLM, served by Cloudflare Workers AI through the native [ai] binding — no external API keys, billed inside the same Cloudflare account the Workers run in). Verified 2026-07-09 via grep -n '@cf/' workers/subagentworkers/src/index.ts workers/subagentdb/src/index.ts: both files declare the identical WORKERS_AI_MODEL constant.

subagentworkers.com

RAG Q&A over Cloudflare's own docs

Retrieval is lexical, not vector: D1 SQLite FTS5, BM25-ranked, over an indexed corpus of 859 pages across 16 Cloudflare products. Generation runs on Workers AI via the native env.AI.run() binding — zero external API keys.

source: workers/subagentworkers/src/index.ts

subagentdb.com

Natural-language-to-SQL over the family's D1 warehouses

The model writes one SQL query per question; the Worker validates it (single SELECT only, every table checked against an allowlist) before it touches a real database, then the same model explains the results in plain English.

source: workers/subagentdb/src/index.ts

Data-cleaning tooling: plugins/cwc-data-engineering

The fleet's other ML surface is upstream of any model: LLM-pretraining-corpus data engineering, packaged as two plugin skills. Quotes below are verbatim excerpts from each skill's own SKILL.md frontmatter description.

`corpus-cleaning-pipeline`

Typed models (Pydantic + FilterReason enum) and verified, profiled performance/library-quality reference notes for an offline LLM-pretraining-corpus cleaning pipeline — schema validation, vectorized quality filters, exact-hash dedup, datasketch-backed MinHash+LSH near-dedup, real tiktoken token counting, a composite quality-ranking score, sharded parquet output.

The pipeline this skill documents is real and profiled — its FilterReason enum members are grepped 1:1 from the pipeline's actual filter-name strings, with a test enforcing that as a live drift gate.

source: plugins/cwc-data-engineering/skills/corpus-cleaning-pipeline/SKILL.md (frontmatter description, read 2026-07-09)

`parquet-shape-profiler`

Compute a structured, typed data-shape profile (DatasetShapeProfile: row/column counts, file size, per-column dtype + kind + null rate, numeric percentiles [min/p25/median/p75/max/mean] or categorical distinct-count + top-values) for any parquet dataset, as validated Pydantic JSON … Every number in the output is a REAL computed statistic, never a placeholder.

Deliberate cross-plugin split, per its own description: this skill computes the numbers; cwc-design's data-shape-svg skill renders the resulting JSON as SVG.

source: plugins/cwc-data-engineering/skills/parquet-shape-profiler/SKILL.md (frontmatter description, read 2026-07-09)

The live writeup of that pipeline work — architecture diagram, sample-shard findings, and the verified near-dedup performance fix — is agenttables.com. Per that site's own scope note, results shown there are sample-data results; the full 40GB corpus run has not happened yet.

Honest scope note. No model training happens in this fleet. No fine-tuning, no embedding pipelines, no vector databases — even the RAG retrieval on subagentworkers.com is lexical (D1 FTS5/BM25), not vector search. What exists is inference (one Workers AI model, two sites) plus offline corpus-cleaning and dataset-profiling tooling. This page inventories that, and nothing more.

agentmachinelearning