OUR APPROACH

Two prerequisites for
a real AI agent.

We believe there are two fundamental prerequisites that must be solved before an AI agent can meaningfully act on your behalf: deep context and reliable capability.

01

Context

Most AI assistants only access data available through APIs — your email, calendar, maybe Slack. But this is a fraction of who you are. The most valuable personal data is locked inside walled gardens that don't offer API access.

Where your data actually lives

Twitter DMs
Walled gardenGUI agent
Instagram DMs
Walled gardenGUI agent
WhatsApp
Walled gardenGUI agent
LinkedIn
Walled gardenGUI agent
Gmail
API accessibleOAuth API
Google Calendar
API accessibleOAuth API
Slack
API accessibleOAuth API
Notion
API accessibleOAuth API

Twitter direct messages, interactions across all platforms, Instagram direct messages — all of this personal data is scattered across platforms. When combined with data that can be granted via APIs (your Gmail, Calendar, Slack), we can finally construct a user profile with the depth for an agent to truly understand who you are, what you want, and what you may need tomorrow or next week.

This allows agents to proactively prepare rather than waiting for you to give a request.

We're not just obtaining data from easily accessible APIs. We're obtaining the entire digital footprint of a user.

For the first time in the internet era, we can aggregate our own data and leverage it to provide value to ourselves — rather than relying on platforms to act nicely and give us this data back. Intelligence has always flowed from users to platforms. We're reversing that flow.

The Personal Intelligence Pipeline

Just like companies build pipelines for business intelligence, we build a personal intelligence pipeline — one that everyone should have and will have. It processes decades of data you've created across all platforms.

Raw Inputs

App connectors, files, conversations

High-fidelity personal data, collected at the source.

SAMPLE OUTPUT
>Subject: "Q2 budget review" — from: CFO
>WhatsApp: "Let's finalize the SF office lease"
>Meeting: Board sync — tomorrow 2pm
>Twitter DM: "Interested in your product"

Indexing

Full-text search, embeddings, vector search

Multi-modal retrieval layer for fast, contextual lookup.

SAMPLE OUTPUT
>FTS index: 124,392 documents across 8 sources
>Vector: [0.234, -0.891, 0.445, ...] dim=1536
>Chunk: "SF office lease" → cluster #47
>Entity: CFO → person_node_0x3f2a

Features

Aggregates, behavioral features, graphs

Derived behavioral signals from raw data.

SAMPLE OUTPUT
>Communication freq: CFO ↔ You → 12x/week
>Topic cluster: "US expansion" — 23 mentions
>Pattern: You research legal topics on Mondays
>Social graph: 3 hops to target investor

Detection & Modeling

Topic models, change points, anomalies

Structure discovery and change detection.

SAMPLE OUTPUT
>Change point: US expansion topic +340% this week
>Anomaly: No reply to legal counsel in 5 days
>Topic model: "corporate setup" → high intent
>Prediction: User will need legal docs by Friday

Insights & Presentation

Dashboards, RAG narratives, nudges

Grounded, explainable outputs delivered as proposal cards.

SAMPLE OUTPUT
>Proposal: Prepare Delaware C-Corp brief
>RAG: 3 relevant lawyers found in your network
>Nudge: Follow up with legal counsel today
>Action: Draft follow-up email — ready to send
02

Capability

A personal assistant agent must be able to operate software interfaces on your behalf. Due to the limitations of APIs, the things agents can do are by nature limited. GUI agents — agents that can interact with graphical interfaces — are always needed to fill the gap. In most cases, they are the only way to get things done.

But the current state of GUI agents is fundamentally broken.

The problem with current GUI agents

Most GUI agents today use a vision-based approach. They take a screenshot, feed it to a vision model, and get back a low-level action to execute:

Left click
Right click
Key press
Scroll

After each granular step, the agent takes a new screenshot and asks the vision model what to do next. This approach is general, but comes with massive trade-offs in cost, latency, and reliability — especially for long-horizon tasks where context explodes and errors compound.

Our approach vs. the status quo

Metric
Vision-based agents
Boxy agents
Send a message
~3 minutes
~3 seconds
Reliability
~60-70%
~100% (deterministic)
Cost per action
$1-5 (VM + tokens)
~$0 (pre-trained)
Approach
Vision model + screenshots
Deterministic manifests
Long-horizon tasks
Context explosion, unreliable
Stable, predictable

Three things that matter

Speed

If you take 30 seconds to find a contact and draft a message, the agent should take 3 seconds — not 3 minutes. Actions must be faster than doing it yourself.

Reliability

When you want to send a message, you need 100% confidence it will be sent with the exact content you consented to. Deterministic actions build trust over time.

Cost

If it costs $3 to send an email via a GUI agent, adoption will be zero. Our pre-trained agents drop the cost to effectively nothing.

We have pre-trained our GUI agents to execute deterministic actions on specific platforms. Instead of guessing what to click next, our agents know exactly what to do — every time. This makes actions fast, reliable, and virtually free. It builds trust so users can gradually delegate more sensitive tasks over time.