Vector Ridge Lab · Paper implementations
Working proofs of concept of recent AI research, reproduced within days of publication and
published as live demos. Research first, production always.
How the Lab is built
An author agent reads AI newsletters every morning, scores each paper for AI engineering relevance
and, when one scores 8/10 or more, implements it as a new tab of the Space. Nothing it writes
reaches production without passing two independent gates.
Author agent (daily, Claude) reads newsletters, scores papers (INBOX.md), writes the POC
│ opens a PR: poc/<slug>-<date> (or inbox/<date> on days without a POC)
▼
smoke (GitHub Actions) imports the app, renders it, serves / and /config, runs every
│ event handler with default and example inputs, checks pins
│ and the website contract → tests/
review (GitHub Actions, Groq) a reviewer that never saw the authoring checks the diff
│ against REVIEW_CHECKLIST.md; reviewer code and checklist
│ come from `deploy`, so a PR cannot change its own rules
▼
auto-merge only when both checks pass; `deploy` is protected, no bypass
▼
Deploy (GitHub Actions) smoke again, upload hf-space/, wait until the Space is RUNNING
▼
vectorridgelabs.com Lab sync rebuilds the catalog from app.py + INBOX.md
Lab store sync (GitHub Actions) mirrors POCs + papers into Supabase for the website's RAG
Weekly numbers for this loop (PRs, reviewer failures, CI blocks, failed deploys) come from
tools/agent_metrics.py and are kept in METRICS.md.
Repository layout
| Path | What it is |
|---|
hf-space/ | The Gradio app published to the Space. One folder per POC, one poc_tab(...) per POC in app.py. |
INBOX.md | Every paper scanned, with score, summary, original source link and implementation status. |
tests/ | Smoke suite and its harness (smoke_harness.py exercises every wired handler). |
REVIEW_CHECKLIST.md | The rules the reviewer applies, each seeded from a real incident. |
tools/kb.py | Lab knowledge base: hybrid retrieval (BM25 + trigram, RRF) over INBOX.md, app.py and the POC folders. |
tools/groq_review.py, tools/check_pr_scope.py | The reviewer and its deterministic scope check. |
tools/agent_metrics.py | Weekly metrics of the agent loop from the GitHub API. |
tools/lab_store_sync.py | Mirrors POCs and scanned papers into the website's retrieval store (Supabase pgvector + full text) after each merge. |
CLAUDE.md, docs/daily-author-prompt.md | Instructions for the author agent (versioned like code). |
hybrid-search/, cad-score/ | The first two POCs as standalone projects, before the Space existed. |
Run it locally
python -m venv .venv && source .venv/bin/activate
pip install -r hf-space/requirements.txt pytest requests
cd hf-space && python app.py # the Space, on http://127.0.0.1:7860
cd .. && pytest tests/ -q # the same gate CI runs
Explore the Lab without opening the big files:
python tools/kb.py brief # tabs, regions, recent scans, backlog
python tools/kb.py search "agent memory cost" # related POCs and papers
python tools/kb.py seen "<paper title>" # already scored or implemented?
Contributing a POC
Open a PR to deploy from a poc/<slug>-<date> branch. Read REVIEW_CHECKLIST.md first: the new
tab is declared with with poc_tab("Name: short claim"):, starts with a gr.Markdown intro, keeps
the pinned gradio==4.44.0, does no heavy work at import time, and links the original paper.
License
Code is released under the MIT License. Each POC reimplements ideas from the paper or
announcement it links; credit for the research belongs to its authors.
