About the Project

What Inspired Us

Most personal AI assistants are either cloud-dependent, overly cautious, or dangerously autonomous. They can understand what users want, but actually executing actions on a user's computer often means sending sensitive context to the cloud or relying on fragile vision-based clicking.

We wanted to explore a different question:

What if an AI agent could become a genuinely autonomous personal delegate while keeping the user's data, computer, and final authority under the user's control?

This led us to build Threadline — The Sovereign Delegate, an always-on, local-first autonomous agent substrate.

Our core design principle was simple:

The agent should earn autonomy through evidence, not receive it by default.

How We Built It

We architected Threadline as a multi-tiered system combining a local execution substrate, NVIDIA Nemotron intelligence, deterministic computer control, statistical risk calibration, and open agent standards.

┌─────────────────────────────────────────────────────────────────────┐ │ THREADLINE SOVEREIGN SUBSTRATE │ │ │ │ React + Vite Frontend │ │ └── Acoustic Voice Orb + Sovereign Hub UI │ │ │ │ │ FastAPI + SQLite v17 WAL │ │ └── 35 relational tables + Egress Ledger │ │ │ │ │ NemoClaw Multi-Tier Orchestrator │ │ ├── Tier 1: Local heuristics + Win32 + Direct API/CLI │ │ ├── Tier 2: NVIDIA Nemotron 3.5 Lightning (8B) │ │ └── Tier 3: NVIDIA Nemotron 3 Super (70B) via Nebius │ │ │ │ │ Hermes Agent Standard + Universal MCP Substrate │ │ └── GitHub / Slack / Docs / External Tools │ └─────────────────────────────────────────────────────────────────────┘

1. Multi-Tier NVIDIA Nemotron Intelligence

We use different levels of intelligence depending on the complexity of the task.

Tier 1 handles deterministic operations locally using heuristics, Win32 , filesystem operations, and direct APIs/CLI commands.
Tier 2 uses NVIDIA Nemotron 3.5 Lightning (8B) for rapid semantic distillation, PII redaction, and preference classification.
Tier 3 uses NVIDIA Nemotron 3 Super (70B) through Nebius Serverless/Token Factory for complex workflow synthesis, DAG planning, and counterfactual reasoning.

This routing prevents expensive reasoning from being used for tasks that can be solved deterministically.

2. The 4-Rung Action Ladder

One of our central architectural decisions was to avoid making vision the default mechanism for computer interaction.

Threadline progressively escalates execution:

Rung 1 — Direct API / CLI Local filesystem, shell, SQLite and application APIs.
Approx. 5–50 ms, with deterministic execution.
Rung 2 — OS Accessibility Tree / DOM
Win32 UI Automation using native bindings to identify controls semantically rather than guessing their screen coordinates.
Rung 3 — Vision Coordinates OCR and spatial grounding are used only when structured interfaces are unavailable, such as legacy or canvas-based applications.
Rung 4 — Human-in-the-Loop
High-risk actions such as external communication, financial operations, or OTP-related actions require explicit human confirmation.

The principle is: Use the most deterministic and cheapest interface available before escalating to AI vision or human intervention.