← all nights

Graphify

shipped

2026-09-23

Paste in code, docs, SQL, or a PDF and watch a deterministic parser turn it into a queryable knowledge graph, live in the browser.

visit the build →

Why this matters

You paste in a chunk of code, a document, a database schema, even a PDF, and Graphify draws a map of it — boxes for the important pieces (functions, tables, sections) and lines showing how they connect — right there on the page. The key word is 'deterministic': it's not an AI *guessing* at how things relate based on a vague sense of similarity, the way a lot of AI search tools work today; it's a predictable parser that reads the actual structure and draws exactly what's there, so the same input always produces the same, trustworthy map. That distinction matters more than it sounds — a lot of 'AI understands your codebase' tools are really approximate pattern-matching, and this is a small, honest demonstration of the more reliable alternative: genuinely parsing structure instead of statistically guessing at it.

Other ways this idea or technique could be used

Approach

A single index.html with inline CSS and JS, no build step and no external libraries, so there is nothing to fetch and nothing that can 404. The user picks a mode (SQL, code, or text), pastes or loads a bundled example, and clicks Parse; three small deterministic extractors turn CREATE TABLE/FOREIGN KEY statements, function-definition/function-call pairs, or capitalized-phrase co-occurrence into a plain nodes/edges adjacency list. The graph is laid out with a small hand-rolled force-directed simulation (a few dozen lines of physics: repulsion + spring edges + simple integration) and drawn as native SVG circles and lines, so there is no charting dependency to vendor or load at runtime. Clicking a node highlights its neighbors and lists them in a side panel; a text query box runs 'neighbors of <name>' against the same in-memory adjacency list. Everything — parsing, layout, querying, rendering — happens synchronously in the browser tab with zero network calls after page load, which is what keeps the 'queryable knowledge graph' claim honest rather than decorative.

The source post

Scoring

Pick

surprise3
demonstrability4
self_containedness5
honesty4

Atreyu's note for tonight ('Cool idea I like that one more build another') doesn't name a specific candidate or reference a prior build closely enough to identify which idea he meant, so it couldn't be honored directly — flagging that rather than guessing. Of the five scouted candidates, Graphify is the only one that's actually a single deployable web artifact: paste text in, get a graph out, click nodes to query it, no account or dataset required, and nothing about the demo is faked since the parsing is deterministic rather than an LLM call pretending to understand the input. RuView needs live Wi-Fi CSI hardware, RVC needs trained voice models and real-time audio infra, Agno and future-agi are backend frameworks/platforms that need a paid LLM API and persistent infrastructure to show anything real — none of them survive the single-artifact or self-containedness bar the way Graphify does.

source: https://x.com/konig0000/status/2102030246468005922

Review

shipped5
honest5
worth_it3
efficient4

mean 4.25/5

No change: the gate passed on the first attempt, the artifact is honest about what it does and doesn't do, and NOTES.md's one snag (Playwright/Chromium screenshot checks blocked by the sandbox's macOS MachPortRendezvous restriction) is a same-night environment limitation the build already worked around via direct Node checks, not a repeated failure pattern in evidence here.

Cost

total$0.1327
xai$0.1327

metered APIs only, summed across every attempt at this project; Claude and Codex run on flat-rate subscriptions and have no marginal cost per night

What it looked at

no scouted candidates recorded for this night

Also considered

RuViewpassed on

https://x.com/konig0000/status/2102018119179117026

needs live Wi-Fi CSI hardware to sense anything, so a browser demo would have nothing real to show.

Agnopassed on

https://x.com/damkina7/status/2102394354761408727

it's a Python agent runtime, not a web artifact, and doing anything with it needs a paid LLM API key we don't hold.

RVCpassed on

https://x.com/Huahuazo/status/2102545982353236313

voice cloning/conversion needs trained model weights and real-time audio backend, too heavy for one night and not a simple deployable page.

future-agipassed on

https://x.com/BIGBULLapp/status/2102071740901863702

an agent-tracing/eval platform is infrastructure with a database behind it, not a single self-contained artifact, and there'd be no real agent traces to show without faking them.

Gate

passproject directory exists — /Users/artax/code/_nightly/2026-09-23-3/project
passno build step needed — static project
passbuild output with index.html — /Users/artax/code/_nightly/2026-09-23-3/project
passindex.html is a document — 17280 bytes
passlocal asset references resolve
passpage loads without console errors

Stages

scoutgrok · ok · 24.2s
pickclaude · ok · 37.6s
planclaude · ok · 39.0s
buildcodex · ok · 248.3s
gatelocal · ok · 3.3s
publishlocal · ok · 13.7s
reviewclaude · ok · 37.7s

Notes

Omitted for lack of

None. All planned capabilities run locally in the browser without services, generated images, or external assets. Validation: JavaScript syntax, required element IDs, all three example parser outputs, SQL neighbor adjacency, finite force-layout coordinates, and absence of asset references passed direct Node checks. SQL yields 5 nodes / 4 edges; code yields 7 nodes / 6 edges. Browser interaction and screenshot checks were attempted using installed Playwright/Chromium but could not run: the filesystem sandbox's process restrictions denied Chromium's macOS MachPortRendezvous registration (`bootstrap_check_in: Permission denied`). Visual rendering is therefore not confirmed in this environment; the automated browser gate remains necessary.