Let’s talk

Selected work

Browse all projects

Agent-assisted decision support

See the full picture
before you decide.

Bring the situation, the evidence and the possible actions into one workspace. Keep your team in control of what happens next.

Explore this project

Private prototype · Demonstration data

Industrial decision workspace
Actual decision-builder prototype showing scenario intake, a causal map and a decision brief.
Guided view of a genuine screen captureFull screen

A closer look at the work, one decision at a time.

Explore the situation, inspect the reasoning and review the decision brief. These are details from the prototype above, with no live equipment connected.

01 / Industrial decision supportPrivate prototype

Industrial Decision Support Workspace

A goal-to-map agent turns the situation into a transparent causal model. Operators can inspect evidence and uncertainty, test scenarios and record the reasoning behind a recommendation.

Read the case study
Actual decision-builder prototype showing scenario intake, a causal map and a decision brief. Enlarge screen
Application capture

Actual prototype screen using a demonstration scenario. Client identifiers are not shown; no live equipment is connected.

Industrial Decision Support Workspace — Decision builder

Actual decision-builder prototype showing scenario intake, a causal map and a decision brief.

Actual prototype screen using a demonstration scenario. Client identifiers are not shown; no live equipment is connected. Open original image in a new tab

The problem

Signals and possible causes were only part of the decision. Operators also needed to weigh operating goals, constraints, procedures, consequences and uncertainty in a form the next shift could inspect.

Where agents fit

A goal-to-map agent builds a stable decision structure, a scenario agent evaluates operating states and a record agent preserves the rationale. Every node remains inspectable and editable.

Human control

The application does not control equipment or make the operating decision. An accountable person does.

Inside the build Workflow, technology, data & hosting

Workflow

  1. State the abnormal condition and operating goal
  2. Connect evidence, causes and constraints
  3. Compare possible actions and consequences
  4. Record the recommendation and final decision

What was built

  • Causal map with confidence and uncertainty
  • Goals, constraints and procedure references
  • Scenario and consequence comparison
  • Decision record with human ownership

Platforms

Browser decision workspace

Technology stack

React 19 · TypeScript · Vite · React Flow · Motion · Lucide

Data and database

Local browser storage · Portable JSON records · No server database in the prototype

Hosting and runtime

Static browser application · Private prototype deployment

Agent responsibilities

Goal-to-map agent. Builds a transparent map of signals, causes, goals, constraints and actions.

Scenario agent. Evaluates telemetry against explicit thresholds and compares consequences.

Decision-record agent. Preserves evidence, confidence, uncertainty and the accountable decision owner.

What to measure

Time to assemble context, missing constraints, scenarios reviewed and decisions with recorded rationale.

These are evaluation measures, not claimed client results.

02 / Scientific workflowPrivate prototype

Scientific Process Control Workspace

We brought setup, simulated run status, events and operator notes into one reviewable workspace. Bounded setup, observer and review agents organize the work, but the prototype does not control laboratory hardware.

Cropped screen from the scientific simulator showing six sample traces and a selected sample detail panel. Enlarge screen
Application capture

Actual simulator capture, cropped to exclude client and product branding. All traces shown are simulated, not experimental results.

Scientific Process Control Workspace — Run-monitoring workspace

Cropped screen from the scientific simulator showing six sample traces and a selected sample detail panel.

Actual simulator capture, cropped to exclude client and product branding. All traces shown are simulated, not experimental results. Open original image in a new tab

The problem

The team needed to see how a future control experience could guide setup, show a run in progress and leave a record that another scientist could review later.

Where agents fit

The setup agent checks the run plan, the observer agent records simulated events and the review agent assembles the history. Scientists keep control of the procedure and any future hardware action.

Human control

This is a simulator-only prototype. It has no functional connection to laboratory hardware.

Inside the build Workflow, technology, data & hosting

Workflow

  1. Prepare the run and review the setup
  2. Start a simulated operating sequence
  3. Watch status, trends and recorded events
  4. Review the history with operator notes

What was built

  • Guided setup and run preparation
  • Simulated status and trend views
  • Event history and operator notes
  • Reviewable run record

Platforms

Browser simulation workspace

Technology stack

React 19 · TypeScript · Vite 8 · Express 5 · Simulation data source · Automated fact checks

Data and database

Simulated run records · In-memory prototype state · No production database

Hosting and runtime

Node.js demonstration server · Private prototype runtime

Agent responsibilities

Setup agent. Checks that the run plan and required setup fields are ready for review.

Observer agent. Tracks simulated status changes and records events without touching hardware.

Review agent. Assembles trends, events and operator notes into a reviewable run record.

What to measure

Setup consistency, review time, incomplete records and readiness for later engineering work.

These are evaluation measures, not claimed client results.

03 / Product configurationConcept product

Vitrifold Product Configurator

A buyer can explore options in the browser, save an image, share a configuration, download a specification and open a prepared inquiry.

Actual Vitrifold configurator showing an amber three-piece luminaire and controls for glass, metal and element count. Enlarge screen
Application capture

Captured from the working Vitrifold configurator. An independent concept product, not a commissioned client project.

Vitrifold Product Configurator — Composition studio

Actual Vitrifold configurator showing an amber three-piece luminaire and controls for glass, metal and element count.

Captured from the working Vitrifold configurator. An independent concept product, not a commissioned client project. Open original image in a new tab

The problem

A buyer could describe the product in broad terms, but the sales team still needed the chosen form, material, finish and dimensions before they could respond with confidence.

Where agents fit

A deterministic configuration agent carries valid choices forward, compiles the specification and prepares the inquiry context. The core workflow does not need generative AI to be useful.

Human control

This independently built concept demonstrates the workflow and is not presented as client work.

Inside the build Workflow, technology, data & hosting

Workflow

  1. Choose the product form
  2. Set material, dimensions and finish
  3. Review the configuration as it changes
  4. Save, share, export or begin an inquiry

What was built

  • Live product rendering
  • Configuration summary and specification
  • Image and PDF export
  • Shareable configuration and prepared inquiry

Platforms

Responsive web application

Technology stack

Next.js 16 · React 19 · TypeScript · SVG rendering · Anime.js · Motion · jsPDF

Data and database

Client-side configuration state · No server database in the concept build

Hosting and runtime

GitHub Pages · Static CDN delivery

Agent responsibilities

Configuration agent. Keeps form, material, finish and dimension choices in a valid state.

Specification agent. Turns the selected options into a readable technical summary and export.

Inquiry agent. Packages the configuration so sales receives a complete starting point.

What to measure

Completed configurations, inquiry completeness, quote-preparation time and option changes before submission.

These are evaluation measures, not claimed client results.

04 / Field planningConcept product

Shoremetric Operating-Window Planner

We connected forecast data with user-set thresholds so field teams can see consecutive daylight hours that meet the operating limits.

Actual Shoremetric planner showing a gust forecast, adjustable operating limits and qualifying work windows. Enlarge screen
Application capture

Captured from the working Shoremetric concept. A point-in-time forecast, not current conditions or operational clearance.

Shoremetric Operating-Window Planner — Forecast and operating limits

Actual Shoremetric planner showing a gust forecast, adjustable operating limits and qualifying work windows.

Captured from the working Shoremetric concept. A point-in-time forecast, not current conditions or operational clearance. Open original image in a new tab

The problem

A forecast alone did not answer the planning question. The team needed to test several conditions at once, find consecutive workable hours and share the assumptions behind the plan.

Where agents fit

A forecast agent retrieves the conditions, while a planning agent applies the operating limits and finds consecutive workable hours. A briefing agent packages the assumptions for the field team.

Human control

Forecasts support planning. Field teams still confirm local conditions and make the operating decision.

Inside the build Workflow, technology, data & hosting

Workflow

  1. Choose the location and planning period
  2. Set the operating thresholds
  3. Compare the forecast with each limit
  4. Share the workable windows and assumptions

What was built

  • Forecast retrieval by location
  • User-set wind, temperature and weather limits
  • Consecutive daylight-window analysis
  • Shareable planning brief

Platforms

Browser planning tool · Live weather-data API

Technology stack

Next.js 16 · React 19 · TypeScript · Weather API integration · Anime.js · Motion · jsPDF

Data and database

Transient forecast API data · Client-side planning state · No server database in the concept build

Hosting and runtime

Static web deployment · External weather-data service

Agent responsibilities

Forecast agent. Retrieves location-specific weather data for the selected planning period.

Window agent. Tests each hour against the job limits and groups consecutive workable periods.

Briefing agent. Creates a shareable plan with thresholds, failures and forecast assumptions.

What to measure

Planning time, workable windows found, threshold failures and changes caused by a new forecast.

These are evaluation measures, not claimed client results.

05 / Supply chain operationsApplication family

Supply Chain Operations Agent

We built a local purchasing agent for Outlook and Excel, plus a voice-first supply chain environment. A manager can ask a spoken question, follow the impact from supplier to customer and review recovery options before approving anything.

Read the case study

The problem

Requests entered through a shared inbox while disruption context lived across supplier, material, production and customer records. The team needed agents that could assemble the picture quickly while preserving the tools, rules and approval steps already in use.

Where agents fit

A realtime voice agent handles the spoken briefing. Routing and recovery agents collect evidence, trace impact and prepare options so a person can make the operational decision.

Human control

The agents cannot award suppliers, place orders, change production, approve substitutions or make customer commitments.

Inside the build Workflow, technology, data & hosting

Workflow

  1. Ask the AI voice agent about a delay or exposed commitment
  2. Trace the effect through materials, production and customers
  3. Compare recovery scenarios and purchasing actions
  4. Approve, edit or reject the action packet

What was built

  • Realtime AI voice briefings with clear AI disclosure
  • Shared-inbox routing and constrained AI triage
  • Supplier-to-customer impact tracing
  • Approval-ready recovery and escalation packets

Platforms

Managed Windows desktops · Browser-based voice cockpit

Technology stack

OpenAI Realtime API · OpenAI Responses API · WebRTC voice · PowerShell 5.1 · Outlook + Excel COM · Windows DPAPI

Data and database

Excel operational records · Local structured files · No shared cloud database in the desktop MVP

Hosting and runtime

Render voice-cockpit service · Managed Windows desktop runtime

Agent responsibilities

Voice briefing agent. Turns a spoken supply question into a traceable operational briefing.

Routing agent. Applies company rules first and reviews only unresolved purchasing requests.

Impact-tracing agent. Connects a supplier or material issue with the production and customer records affected.

Recovery agent. Compares scenarios and assembles an approval-ready action packet.

What to measure

Time to trace an exposure, request-assignment time, overdue work, recovery options reviewed and manual rework.

These are evaluation measures, not claimed client results.

Client names and account details are withheld. No reconstructed interface is shown.

06 / Sales operationsInternal application

Sales Opportunity Agent

The native Mac application uses research, qualification and drafting agents to prepare an editable opportunity, show the evidence, flag missing questions and draft the next email for review.

Read the case study

The problem

Useful context arrived in different formats and at different stages of a sale. Staff needed one place to understand the account, test the fit, see what was missing and prepare a relevant next step.

Where agents fit

Research, qualification and drafting agents share one evidence record. They can prepare a recommendation and a relevant email, but the seller decides whether the opportunity fits and what leaves the company.

Human control

The application prepares and saves drafts. A person reviews the evidence and decides what to send.

Inside the build Workflow, technology, data & hosting

Workflow

  1. Import an email, document or clipboard note
  2. Edit the intake before creating an opportunity
  3. Compare the evidence with configurable services
  4. Review the gaps, risks and next email

What was built

  • Editable intake before anything enters the queue
  • Configurable service and qualification profiles
  • Evidence, fit risks and missing questions
  • Drafts for each stage of the sales process

Platforms

Native macOS application · Local document and clipboard intake

Technology stack

SwiftUI · Swift · OpenAI Responses API · OpenAI web search · macOS Keychain · Foundation file storage

Data and database

Local JSON opportunity record · Evidence and drafts stored on device · No server database

Hosting and runtime

Native macOS runtime · OpenAI hosted model services

Agent responsibilities

Research agent. Finds company context and keeps source evidence attached to the opportunity.

Qualification agent. Tests the evidence against configurable services and preserves uncertainty.

Drafting agent. Prepares stage-specific follow-up that a seller can edit, copy or save.

What to measure

Preparation time, follow-up turnaround, qualification gaps and opportunities that still need research.

These are evaluation measures, not claimed client results.

Client names and account details are withheld. No reconstructed interface is shown.

Start with your business

Your project may
look nothing like these.

That’s fine. Bring the process that keeps breaking, the tools that don’t connect or the application your team wishes it had. We’ll work out what is worth building—and what isn’t.

Talk through your project