Intent Canvas

A browser-first spatial workspace for turning business intent into inspectable computation. Place evidence, connect ideas, describe what you want — get back a plan you can trust, then execute.

Showcase

Section 1 — Intent as Input	Section 2 — Spatial Canvas
Stop clicking buttons. Express intent.	Place evidence spatially. Let proximity become meaning.

Section 3 — Inspectable Plan	Section 4 — Grounded Results
Every step is explainable before it runs.	Every output is evidence-bound and validated.

Screenshots are from the live canvas at 1280px — mobile reflow is verified at 390px with no overflow.

Why Intent Canvas

Most tools force you to click through forms. Intent Canvas lets you think spatially — lay out datasets, documents, notes, and examples, connect what relates, and describe the outcome in plain language. The system compiles your spatial thinking into a plan you can inspect, not a black box.

Product model

Context nodes + spatial relationships + user intent
                         |
                         v
                  SpatialGraphAST
                         |
                         v
                   Inspectable plan
                         |
                         v
                  User approval
                         |
                         v
                 Results on canvas

What you can do

1. Place datasets, documents, notes, examples on the canvas
2. Connect with explicit edges; proximity (240px default) forms spatial clusters
3. Describe the outcome you want
4. Inspect the plan — sources, confidence, workflow, verification
5. Approve — plan is bound to this canvas state, runs once
6. Receive results as nodes you can keep, move, or save as a primitive

Workspace features

• Spatial — pointer + keyboard, pan/zoom, drag, file drop, relationships
• Nodes — dataset, document, instruction, example, output, custom primitive; PDF/CSV/TXT/MD/JSON/raster uploads (SVG rejected)
• Plan review — context rationale, workflow stages, confidence, assumptions, constraints
• Execution — 4 capabilities: data patterns (safe SVG), document synthesis, meeting insights, UI concepts
• State — Zustand persists only bounded workspace data; transient plans/results are not persisted
• Primitives — compile your graph into a reusable custom primitive

Capabilities

Tool	Needs	Returns
DataPatternFinder	Dataset	Observations, anomalies, safe SVG chart
DocumentSynthesizer	Document / dataset / instruction / example	Takeaways, connections, contradictions
MeetingInsightExtractor	Document / instruction	Decisions, action items, risks
UIConceptGenerator	Example + explicit UI intent	Hierarchy, styling, theme, referenceBasis

Planner chooses the smallest useful set. No tool runs because its input exists.

Quick start

npm install
cp .env.example .env   # add VITE_GEMINI_API_KEY
npm run dev            # → http://localhost:3000

Paste your Gemini API key in the top bar (or set VITE_GEMINI_API_KEY). Get one at https://aistudio.google.com/app/apikey — stored in localStorage, never sent elsewhere except generativelanguage.googleapis.com.

Configuration

VITE_GEMINI_API_KEY=AIza...
VITE_GEMINI_MODEL=gemini-2.5-flash
VITE_CANVAS_ID=workspace_canvas
VITE_SPATIAL_CLUSTER_ID=primary
VITE_PROXIMITY_DISTANCE_PX=240
VITE_DEFAULT_INTENT_PROMPT=Analyze the supplied context, identify the most useful supported outcome, and propose grounded next steps.
VITE_PRIMITIVE_TITLE=Evidence-Based Risk & Opportunity Primitive
VITE_PRIMITIVE_DESCRIPTION=User-composed dynamic computational primitive

Limits — intentional bounds

• Upload 10 MB, PDF text 10k chars, image preview bounded, SVG rejected
• Workspace: 30 nodes, 60 edges, output 100 KB
• Inline SVG sanitized and bounded
• Browser persistence is per-profile, not multi-user

Scripts

npm run dev      # Vite dev
npm run build    # tsc + vite build → dist/
npm run preview  # serve dist

Deploy dist/ to any static host.

Built for the NYC Codex Community Hackathon — where spatial thinking becomes an inspectable plan.

---

Built by Drishtant Ghosh (Drix10) · AI systems engineer and 1x acquired founder, Bengaluru

Portfolio · LinkedIn · X · Blog · All projects

_{More from Drix10: Agent Flow (guardrails for AI coding agents) · MiroHedge (AI-assisted systematic fund) · Guardian (UPI scam interceptor) · Sentinel (AI security CLI) · Keystroke-LLM (NumPy transformer on a keyboard) · AI Resources (daily AI knowledge base)}