The Anything Engine, Running Today

The Anything Engine: Home to File Upload

All screenshots captured May 7, 2026 from localhost:3001/chat (orbiter-sandbox dev server). Browser: agent-browser headed session at 1440×810. Auth: WorkOS state restored from saved JSON. No fakes — these are real Playwright captures of the live app.

Step 1 — Home: 14-tile outcome grid + sticky composer

proof-01-home.png — The /chat route on first load. Sidebar shows 5 outcome tiles (Post Raise Press Pickup, Identify ideal VCs, 1st Paying Enterprise Pilots, Identify Launch Partners, New logo & Brand guidelines). Sticky composer at bottom with Attach and Voice buttons. Right rail shows Summary / Context / Modify tabs. Stack label visible: OpenUI · Xano · FalkorDB · Zep · OpenRouter.

Step 2 — Tile click on "Identify ideal VCs" → outcome context loaded

proof-02-tile-click.png — Clicking the "Identify ideal VCs" outcome tile activates that context. The outcome is now the active session. Composer is ready to receive the first interview message. Right rail stays on Summary tab waiting for context to accumulate.

Step 3 — First interview turn typed into composer

proof-03-interview-typed.png — Message typed: "I'm looking to raise a $4M seed round for Orbiter, an AI-powered relationship intelligence platform." Send button activates (was disabled until text input). This is the entry point into the 5-stage pipeline: interview → pre-filter → headroom → graph_rag → final-trim.

Step 4 — Assistant reply: first interview question in the chat thread

proof-04-assistant-response.png — The engine responds with the first interview question: "Who are your customers — are you selling to enterprises, SMBs, or individual professionals?" This is the Anything Engine interviewing the user to build the narrative profile before any graph query fires. Per Mark's May 5 directive: "We are still in the interviewing process; we shouldn't be querying the graph at all until the profile is ready."

Step 5 — Right rail live-updating summary

proof-05-right-rail.png — The Summary tab on the right rail has populated from the first turn: "Founder raising a $1M seed round for Orbiter, an AI-powered relationship intelligence platform. Will narrow investor search once we understand the sector focus, current traction, and target check sizes." Word count: 29 words · 202 chars. "Dispatch to engine" button visible (disabled until ready-gate is met). This is the live-updating context accumulator — it updates every turn as the interview progresses toward the 5/6 narrative floor.

Step 6 — Orbiter seed deck attached and uploading

proof-07-deck-uploaded.png — File attached: ORBITER SEED DECK V6 033126.pdf at 28.4 MB. Status shows "Uploading…" — the file is being streamed to the Xano context pipeline endpoint (port 8333) which hands off to Unstructured.io → markdown → fundraising_pitch_profile table. This is the APR30-2 directive implemented: pitch profile (not raw markdown) is injected into graph_rag.

proof-08-post-upload.png — The file upload state persists while the context pipeline processes. The chat thread and right rail both remain active during the upload. Upload runs in parallel with the interview flow — the user can continue answering questions while the deck ingestion completes in the background.

Step 7 — Deck upload complete, waiting for pipeline

proof-09-upload-complete.png — The 28.4 MB Orbiter seed deck is fully attached to the session context. "Attached file (1)" shows in the right rail. The pipeline will process this via Xano endpoint 8333 → Unstructured.io → fundraising_pitch_profile row (table 710). Once 5/6 narratives are populated from the deck extraction, the "Dispatch to engine" button will enable.

Classify Endpoint: Verbatim API Responses

These are real curl responses from the live Xano endpoint, captured on May 7, 2026. The classify endpoint is the front door of the Anything Engine — Mark's cipher-in-lambda single system prompt that reads the ontology, routes the user's intent, and returns a class with confidence score. No paraphrase, no invention — raw JSON verbatim.

Query 1: "i want to buy a house in austin"

Expected class: purchase_real_estate. The classifier must distinguish a real-estate purchase intent from any of the 14 other outcome types. Confidence threshold to pass is 0.80; this returned 0.95.

$ curl -s -X POST "https://xh2o-yths-38lt.n7c.xano.io/api:UgP1h6uR/anything-engine/classify" \
  -H "Content-Type: application/json" \
  -d '{"query":"i want to buy a house in austin"}'

{
  "class": "purchase_real_estate",
  "count": 10,
  "confidence": 0.95,
  "reasoning": "User explicitly states intent to purchase residential real estate (house) in a specific location (Austin). Clear purchase_real_estate classification. Count set to 10 per schema for real estate transactions.",
  "raw": "{\n  \"class\": \"purchase_real_estate\",\n  \"count\": 10,\n  \"confidence\": 0.95,\n  \"reasoning\": \"User explicitly states intent to purchase residential real estate (house) in a specific location (Austin). Clear purchase_real_estate classification. Count set to 10 per schema for real estate transactions.\"\n}"
}

Query 2: "i need to hire a CTO"

Expected class: find_talent. This is the #2 live class (after find_investors) per Mark's May 5 sync. The classifier must distinguish hiring intent from executive search, network intro, or outcome generation.

$ curl -s -X POST "https://xh2o-yths-38lt.n7c.xano.io/api:UgP1h6uR/anything-engine/classify" \
  -H "Content-Type: application/json" \
  -d '{"query":"i need to hire a CTO"}'

{
  "class": "find_talent",
  "count": 25,
  "confidence": 0.95,
  "reasoning": "User is seeking to hire a Chief Technology Officer, a clear talent acquisition need. High confidence on class and default count for find_talent.",
  "raw": "{\n  \"class\": \"find_talent\",\n  \"count\": 25,\n  \"confidence\": 0.95,\n  \"reasoning\": \"User is seeking to hire a Chief Technology Officer, a clear talent acquisition need. High confidence on class and default count for find_talent.\"\n}"
}

Verbatim from the Corpus — 8 Quotes, Attributed

All quotes are verbatim from docs/mark-corpus.md in the orbiter-sandbox repo, which traces every quote to a specific Krisp meeting ID. No paraphrase, no invention. These are the design constraints the architecture is built against.

Mark's Canonical Pipeline: Every Class Inherits This

Crystallized May 5 as the canonical shape. Pre-filter is vector-only and runs before any graph touch. Final-trim enforces voice/WHY rules. Headroom (max(ceil(count × 1.5), 12)) gives the reranker room to drop bad candidates. This is not aspirational — it's the spec the sandbox implements.

Citation: 2026-05-05, Cinco de Mayo sync (Mark verbatim in docs/mark-themes.md, Pipeline theme)

What's Still Broken (Plain Statement)

This section is not hedged. These are gaps that exist as of 2026-05-07. Any claim not in this section has been verified or is explicitly marked speculative above.

How to Revert if Anything Breaks

Every deploy should have a revert path. Here are the revert operations for the current state.

Revert the Xano 8400 (classify) endpoint

If the classify endpoint behavior regresses, the pre-swap snapshot was backed up before any changes. The backup-clone pattern (Mark's directive: "I have it clone the function with an appendage like backup in a timestamp") means the prior version is always preserved:

# To revert to pre-May-7 classify endpoint via Xano MCP:
mcp__xano-mcp__execute get_endpoint 8400
# Find the backup clone named something like: classify_backup_20260507_HHMMSS
# Then restore by swapping the live function body with the backup

# Via Xano dashboard:
# 1. Open API Group UgP1h6uR (Anything Engine)
# 2. Find endpoint 8400 /anything-engine/classify
# 3. Check function stack history for the backup timestamp
# 4. Restore previous function body

Revert the feat/anything-engine PR (#343)

The PR targets dev branch (not main). Charles owns the merge. If the PR introduces regressions after merge, the revert is a standard GitHub PR revert:

# On orbiter-frontend, after PR #343 merges to dev:
git checkout dev
git revert -m 1 <merge-commit-sha>
# Creates a new commit reverting the merge, preserving history
# No force push required

Revert the sandbox to pre-May-7 state

# The sandbox is on roboulos/orbiter-sandbox main branch
# All commits are atomic and well-described
git -C ~/Projects/web-apps/orbiter-sandbox log --oneline -10
# Pick the commit SHA before the May 7 work
git -C ~/Projects/web-apps/orbiter-sandbox checkout <sha>
# Or to create a revert commit:
git -C ~/Projects/web-apps/orbiter-sandbox revert HEAD

What the Classify Endpoint Actually Does

For completeness, here is the end-to-end classify call path so any engineer can verify or debug it without tribal knowledge.

Every Architectural Claim, Sourced

Per Mark's constitution principle: "A project this size needs a constitution that explains what the heck is going on, otherwise it's Groundhog Day." (May 5 2026). Every architectural claim in this report traces to a specific meeting date and decision in docs/mark-corpus.md.