StageLoop

Give it a listing URL and $5. StageLoop scrapes a real apartment, rebuilds it in 3D, stages your real furniture, critiques its own work, and loops until it's right — paying its own bills via x402.

Comment

Inspiration

Moving is a leap of faith: you sign a lease without knowing if your bed even fits. We'd built NestFit-style 3D staging before this hackathon asked a harder question: can an agent run the whole pipeline itself, pay for its own data, judge its own work, and iterate until it's right?

What it does

Give StageLoop your furniture (real dimensions) and a live listing URL. It then runs a fully self-directing loop:

  • FETCH the agent pays for a live scrape of the listing (Zero x402 micropayment to a Browserbase capability) and downloads the real photos. Real tx on our demo: 0.01 USDC, receipt + txHash archived per run.
  • VISION every photo is analyzed by a vision LLM on AkashML (rooms, windows, doors).
  • PLAN it synthesizes a true-scale, room-by-room floor plan.
  • ACT an LLM interior designer places your furniture. v1 is intentionally raw.
  • OBSERVE a VLM critic plus a deterministic geometric audit (overlaps, blocked doors, beds outside bedrooms) scores it /10.
  • SELF-CORRECT the critic's issues feed back verbatim into the next placement round. Loop until score ≥ 7 with zero geometric faults (max 3 variants).

Every variant yields a 2D plan, a photoreal Blender render, and a walkable 3D scene (dollhouse + first-person walk with collision). Every run archives its goal, timeline, payment receipts and renders — so the improvement arc (3/10 → 3/10 → 8/10 accepted on a real Craigslist listing) is fully auditable.

Sponsor tools all load-bearing, none decorative

  • Zero (zero.xyz): the agent has its own wallet and pays per-scrape via x402 (wallet 4.9655 → 4.9553 USDC across demo runs; receipts in every run folder). Bonus: hello-world site hosted on Zero at sites.withzero.ai/nestfit-hello-loop
  • Akash: ALL inference (Llama-3.3-70B placement + Qwen3 VLM vision/critique) runs on AkashML, and the entire app FastAPI + React + headless Blender is deployed on an Akash provider (DSEQ 1784326132472, $1.73/mo).
  • Nexla: listing data normalization runs through a live Nexla Express flow (webhook source 125754 → hosted Python transform nexset "stageloop-normalized"), provenance recorded per run.

How we built it

Python/FastAPI agent core; React + three.js frontend (walk/dollhouse modes, GLB scene contract); Blender headless Cycles for photoreal renders; PIL for deterministic plans; Docker (amd64) → GHCR → Akash SDL deployment.

Challenges we ran into

Reasoning VLMs spending their whole token budget thinking (fixed with headroom); an over-strict critic punishing the fixed inventory (prompt ground rules + separating deterministic geometry audit from taste); merge bugs between LLM placements and heuristic fills; arm64 images exec-format-erroring on Akash; expired agent session tokens silently degrading paid fetches (caught by our own run telemetry, fixed live).

Accomplishments we're proud of

A real paid scrape → 3D walkthrough pipeline with an auditable self-correction arc, deployed end-to-end on decentralized compute, built in one day.

What's next

GPU renders on Akash, richer furniture libraries, and multi-listing comparison shopping the agent pays for several scrapes and tells you which apartment actually fits your life.

Built With

Share this project:

Updates