The Data Life
ShippedJun 2026 – Oct 2026Solo builder

Frontend Oracle

A knowledge-graph clone of our tech lead's code-review lens — 239 evidenced edges wired to Claude — plus a VHS mockumentary to launch it.

Source ↗
next.jsclaudesqlitepythoncomfy-cloudhyperframes
VHS frame: two unbothered senior programmers at a desk flanked by chrome cyborgs, under a yellow title band reading OPERATION: BOTTLE THE ORACLE, with a DRAMATIZATION chyron.

Overview

Every engineering team has the same single point of failure: the one or two people who know why everything works. On ComfyUI_frontend that's our tech lead — and 'ask the senior' doesn't scale, doesn't survive vacations, and gates every review on one person's waking hours.

The Frontend Oracle bottles the lens instead. A signal-first pipeline ingested six months of his documented judgment — 10,909 Slack messages, 1,740 PRs, 1,770 inline review comments, 1,047 review verdicts, with bot-generated noise flagged and excluded — and distilled it into a knowledge graph: 50 nodes, 239 edges, 688 citations, zero edges without evidence.

The graph serves three ways: an interactive explorer, an Ask-the-Oracle app where Claude composes advisory answers through the documented lens with citations to the exact Slack thread or review verdict, and an MCP server that grounds code review in Claude Code — so a PR gets checked against the team's own principles instead of vibes.

It shipped with a launch film: The Oracle Tapes, a 2:16 VHS true-crime mockumentary about cyborgs extracting the knowledge, generated with Seedance 2.5 on Comfy Cloud and assembled in HyperFrames. Every CRT screen in the film shows real app output — and a post-credits stinger announces the Oracle's CLI.

Media

The Oracle Tapes, 60-second cut: tape insert → extraction → the Oracle citing its sources, in one minute. Sound on for the modem solo.
The full tape (2:16): a 1994 VHS true-crime dramatization of bottling the oracle. Every CRT screen is real app output. Stay past the credits — the stinger announces the CLI.
A real answer to a real question — relevance gate, documented lens, and two genuine Slack citations.
Extraction complete: the real telemetry subgraph, drawn with the pipeline's true numbers.
The whole architecture as a 1994 mainframe diagram: retrieval in yellow, receipts required.
The correct reaction to grounded retrieval.
The two-up character sheet that kept both oracles consistent across 11 generated shots.
Fine print rolls like it's 1994; the last line is a review verdict.

Features

Ask the Oracle

Type a planning or review question; retrieval ranks the graph's edges, then Claude answers through the documented lens — with a relevance gate (DIRECT vs EXTRAPOLATION) and inline citations. Runs retrieval-only without an API key.

Graph explorer

The full 50-node, 239-edge graph as an interactive map: principles, concerns, positions, and the evidence behind every edge.

Oracle-grounded code review

A stdio MCP server exposes consult_oracle, list_principles, and search_evidence to Claude Code, so reviews cite the team's own rubric — 12 principles, 9 concerns — with primary sources.

Receipts everywhere

688 citations with Slack permalinks and PR links. Every one of the 239 edges carries evidence; the graph is auditable, not a vibe.

The Oracle Tapes

A 2:16 launch film plus a 60-second cutdown: 1994 VHS dramatization, deadpan true-crime narration, a demo where the terminal literally cites its sources, and a post-credits stinger announcing the CLI.

Technical

Pipeline: Python ingesters pull GitHub (PRs, review verdicts, inline comments) and Slack into a SQLite corpus with signal-first bot flagging — 143 authored and 287 reviewed bot PRs excluded — then curated Markdown exports feed a knowledge-graph build that emits kg_graph.json and a self-contained explorer. One script refreshes the app from the graph.

Serving: a Next.js app behind org-only SSO. Retrieval tokenizes the question, scores edges, and hands the top 16 edges plus evidence to Claude, which composes an advisory third-person answer with citations. The same retrieval runs keyless as pure quote-ranking, and as an MCP server inside Claude Code.

The film: HyperFrames renders video from HTML. 11 Seedance 2.5 shots (draft passes, then 1080p finals) hard-cut on one track; the CRT inserts are hand-built HTML sub-compositions rendering genuine API responses; narration is a consented voice clone; the ACE-Step score is EQ-carved under the VO; a global VHS pass adds scanlines, grain, tracking wobble, and a deterministic burn-in clock.

Evidence or it doesn't ship

The graph build refuses edges without citations — 0 of 239 edges lack evidence. That one constraint is what makes the Oracle trustworthy enough to consult during real reviews.

Signal-first ingestion

Bot and agent-generated content is flagged at ingest, not filtered at query time, so the lens stays the human's and the corpus counts stay honest.

A model of judgment, not a person

The persona is advisory third-person by design ('The Oracle's lens suggests…') and declines impersonation. The clone handles the quick questions; the humans remain for novel catastrophes.

The demo is the test

Every screen in the film renders real retrieval output — the citations on the CRT are genuine Slack quotes — so the launch film doubles as an honest product walkthrough.

Social posts

LinkedIn

Draft
M
Miles

We had a problem every engineering team has: two people who know why everything works, and a review queue that only moves when they're awake.

So we cloned one of them. Sort of.

Six months of Slack (10,909 messages) and 1,740 PR reviews, distilled into a knowledge graph of our tech lead's engineering judgment: 50 nodes, 239 edges, 688 citations — zero edges without evidence. Wired to Claude. Ask the Frontend Oracle a planning or review question and it answers through his lens, citing the exact Slack thread or review verdict it's standing on.

It's RAG, but the documents are opinions.

The point isn't replacing the humans. It's that "ask the senior" shouldn't be a single point of failure. The clone handles the quick questions; the humans remain for the novel catastrophes.

And because no serious infrastructure project is complete without a deeply unserious demo film, the explainer is a 1994 VHS true-crime dramatization: cyborgs extracting knowledge from two unbothered senior programmers, junior devs crowding a CRT, a narrator who never winks.

How the tape was made:
• Shots: Seedance 2.5 drafts → 1080p finals on Comfy Cloud
• Character + scene references: Nano Banana
• Narration: a consented voice clone of me (Chatterbox)
• Edit, overlays, VHS grade: HyperFrames — the whole film is HTML
• Every CRT insert shows real Oracle output; the on-screen citations are genuine

🔊 Sound on. The modem solo deserves it. And stay past the credits.

#AIEngineering #KnowledgeGraph #RAG #ComfyUI

X (thread)

Draft
M
Miles
1/3

We cloned our tech lead.

Legally. Sort of. 6 months of Slack + 1,740 PR reviews became a knowledge graph of his engineering judgment — 50 nodes, 239 edges, 688 citations — wired to Claude. It answers like him. With receipts.

Then we made the training video. On VHS. 🔊

269 / 280
M
Miles
2/3

📎 The Frontend Oracle is RAG where the documents are opinions. Ask a planning question, it retrieves the principles our seniors actually wrote down — Slack takes, review verdicts — and answers with citations. 0 of its 239 edges are missing evidence
0 / 280
M
Miles
3/3

Director's cut (2:16). The tape: Seedance 2.5 shots on Comfy Cloud, Nano Banana character refs, a voice-clone narrator (Chatterbox), cut in HyperFrames — the whole film is HTML. Every CRT screen is real Oracle output. Stay past the credits.

240 / 280