Skip to content
eidosworksGet a Friction Review

Intelligent Systems

Give assistance a job. Keep people in control.

We develop AI assistants, published-source retrieval, and React applications around specific tasks. Explore how each system works.

See the projects ↓

Selected work

A useful answer needs an inspectable boundary.

Wellway supports personal reflection. Ask Eidos answers questions about published work. Sentinel Lab is a place to test research ideas. Each has a different purpose and a different level of evidence.

Reflection tool · fictional scenario
Wellway fictional member workspace with a check-in, journey chart, and editable plan.
Existing public capture · fictional scenario data

Reflection tool · fictional scenario

Wellway

The problem
Personal notes and check-ins can be hard to revisit when they live in separate places.
How it works
A guided workspace brings check-ins, a journey view, an editable plan, and optional assistance together around the member’s own reflection.
Human control
The person chooses what to record and review. The advisor view is part of the fictional demo; AI does not approve a plan. Local prompts remain available when hosted assistance is unavailable.
What is verified
The rebuilt demo and 23 local data, evidence, and storage tests passed. The public scenario uses fictional people and records. This is a reflection tool, not diagnosis, treatment, or emergency care.

Behind the build

React web application development & interactive data visualization

A separate member-and-advisor demonstration with editable fictional records, a journey chart, and browser-local workspace storage.

Demo application
  • React
  • TypeScript
  • Vite
  • SVG charts
  • Browser localStorage
  1. Views backed by a shared typed workspace

    React components organize check-ins, plans, reflection notes, and the advisor view. TypeScript data definitions keep the fictional workspace consistent as records change.

  2. Interactive SVG charts and local persistence

    SVG renders the journey chart and selected data points. Browser localStorage saves the demo workspace; validated backup import, export, and recovery handling support reopening without silently discarding a damaged save.

  3. Optional assistance with a visible fallback

    The interface checks assistance availability and keeps local prompts usable when the hosted service is unavailable. The public scenario demonstrates reflection and application design using fictional people and records.

Explore the fictional demo
Published-source assistant
Diagram of the source-answer flow · no provider request shown

Published-source assistant

Ask Eidos

The problem
Visitors need a specific answer about studio work without guessing which page to open.
How it works
The assistant uses published studio information and shows links to its sources. A separate optional AI elaboration is labeled when available.
Human control
Visitors choose what to ask, inspect source links, and close the assistant. Optional AI sends the question to a provider; visitors are asked to avoid private information.
What is verified
The source-answer exercise below uses three published project examples and runs without an AI request. The showcase checks cover question selection and visible source links. The current hosted model and paired provider behavior remain unverified.

Behind the build

AI assistant development & published-source retrieval

A React chat interface backed by a Cloudflare Pages Function, with published-source answers and a separately requested optional AI follow-up.

Implemented source
  • React
  • TypeScript
  • Cloudflare Pages Functions
  • OpenAI Responses API (optional)
  1. Source selection and answer links

    The default server path selects from approved published studio information and returns an answer with source links. The source-answer exercise on this page uses three fixed public examples without a model request.

  2. A separate server-side AI integration

    An explicitly requested enhanced answer can use the OpenAI Responses API when provider bindings and feature flags are configured. Request validation, rate limits, and a daily token budget are implemented server-side.

  3. A bounded assistant, with fallback behavior

    Questions remain in the current tab. Source-based answers remain available when enhanced assistance is disabled or unavailable. Source inspection establishes the implementation; it does not verify the current production model or provider settings.

Try a source-backed answer
Research-stage system
Sentinel Lab research interface showing experiment and evidence status.
Existing public research-interface capture · current status lives in the Lab

Research-stage system

Sentinel Lab

The problem
Experiments can look conclusive before their inputs, failure cases, and evidence have been examined.
How it works
A separate research application makes experiment status and evidence gates visible for streaming prediction and surprise investigations.
Human control
A person reviews experiment status and evidence before drawing conclusions. The public assistant does not operate Lab experiments or take action on a visitor’s behalf.
What is verified
The interface and engineering checks are available for inspection. Synthetic tests do not establish scientific qualification, field performance, or a customer security outcome. Research validation remains open.
Inspect the research boundary

Try a supported task

Inspect the answer and its sources.

This demonstration uses three published project examples. No AI request is made.

Published-source answer

Promo Photo Organizer is a private Data Graphics photo-intake workflow. Job details, labeled photo views, and an upload queue keep context with each image before handoff review. Customer photos stay private. Automatic downstream sorting remains unverified.

Fictional health records in Wellway are demonstration data. Private employee records and customer photos are never part of this example.

How we work

Define the task, then test the limit.

  1. 01

    Name the decision

    Start with the question a person needs answered and the approved sources it can use.

  2. 02

    Set the controls

    Decide who may see, change, or approve each result; make cost and permission limits explicit.

  3. 03

    Check failures

    Test missing sources, uncertain answers, unavailable assistance, and recovery before broadening access.

Starting point

What should AI help with — and what should it never control?

Bring the task, the information involved, and the boundary. We’ll help determine whether AI belongs in the workflow at all and what a useful implementation would need.

Get a Friction Review →Start a Project ↗

A little studio intelligence

Hello. I’m Eidos.

Ask about the work, your next website, or where an idea could go. I start with the studio’s published information.

0/900

This conversation stays in this tab. An optional AI follow-up sends your question to our AI provider; avoid private information. I cannot quote a project or operate the Lab. Details

Want a public conversation? →