CURSIVIS-1.02

Select anything. Press a button. Airia handles the rest.

Comment

Inspiration

Every enterprise AI tool I've used follows the same friction loop:

  1. Open a chat window
  2. Explain the context
  3. Paste the content
  4. Wait for a response
  5. Manually copy the result back

That’s 5 steps of overhead per interaction.

For someone triaging emails, reviewing contracts, or solving MCQs at scale — this compounds brutally.

Quantifying the Problem

Let:

  • t_overhead ≈ 45 seconds
  • n = 30 tasks/day

T_lost = n × t_overhead = 30 × 45 = 1350 seconds ≈ 22.5 min/day

For a team:

  • m = 50 users

T_team = m × T_lost = 50 × 22.5 = 1125 min/day ≈ 18.75 hours/day

Core Insight

The real problem is not AI capability.
The problem is interaction overhead.

Goal:

t_overhead → 0

That became Cursivis Enterprise.

What It Does

Cursivis turns your cursor into an Airia-powered enterprise agent.

Interaction Model

  • Select → highlight anything
  • Trigger → invoke AI instantly
  • Review → get contextual output
  • Act → execute if needed

Interaction cost ≈ O(1)

No chat window. No context switching. No copy-paste.

Context-Aware Intelligence

Same trigger, different behavior:

  • Email → summarize / reply
  • Contract → explain
  • Support ticket → triage
  • MCQs → batch answer

f(selection) → optimal action

System Architecture

Selection → Companion → Backend → Airia Pipeline → Result → Execution

Airia Agent Backend (Node.js)

Single abstraction layer for all AI calls:

POST https://api.airia.ai/v2/PipelineExecution/{pipelineId}

Behavior = f(pipeline configuration)

No code changes needed — only pipeline updates.

Windows Companion App (WPF / .NET 8)

  • Captures text + screen
  • Floating overlay UI
  • Voice + trigger input
  • Routes to backend

Modes:

  • Smart → automatic
  • Guided → suggestions + choice

Browser Execution Layer

AI Output → Structured Plan → Browser Execution

  • Runs inside real logged-in browser
  • No sandbox
  • No re-auth

This is what makes it an agent.

The Airia Integration

Single entry point: airiaClient.js

Prompt Abstraction

[Cursivis Enterprise — legal role]
Action: explain clause
Content type: contract

Selected content:
{selected text}

Why This Works

Context = structured prompt + role injection

  • One pipeline
  • Multiple roles (legal, support, sales, ops)
  • Zero code branching

Each response includes executionId → full traceability

Challenges

1. Schema Constraint

Airia enforces:

{ intent, actions[] }

Fix: encode everything inside allowed fields (output, description)

2. MCQ Bulk Answering

Problem: fewer answers than questions

Fix:

actions[i] = { label: Qi, output: ai } for i = 1...n

n = number of detected questions

3. Clipboard Race Condition

Problem: clipboard delay

Fix: 180ms retry

4. Browser Planner Bug

Problem: wrong function structure

Fix: correct wrapper

What I Learned

  • Pipelines > hardcoded logic
  • Constraints → reliability
  • UX defines perceived intelligence
  • Removing friction = real agent feel

Built With

  • Airia — pipeline execution, routing, planning
  • Node.js + Express — backend
  • WPF / .NET 8 — companion app
  • Playwright — fallback automation
  • Chromium Extension (MV3) — browser execution
  • C# Named Pipe IPC — communication

Built With

Share this project:

Updates