Productivity / Wellness 0→1 Product / MVP

Comigo

AI Companion for ADHD Task Planning & Executive Function Support

~2 mo

Time to MVP

Web + Ext

Surfaces

Streaming

AI Responses

Stripe

Payments

Executive Summary

We designed and built a memory-aware AI companion that helps people with ADHD bridge executive-function gaps — translating scattered thoughts, emotional friction, and shifting priorities into structured, actionable plans. Comigo combines therapeutic-style conversational support with execution-oriented task assistance, using a LangGraph agent workflow and an auto-compacting memory layer to stay useful across repeated interactions without becoming slow or expensive.

The Problem

Comigo's core challenge was not generating task suggestions — it was maintaining useful continuity across conversations while adapting to constantly changing user needs. The system had to store and reuse long-term conversational memory so the agent could recognize recurring struggles and behavioral patterns, support frequent task reprioritization rather than treating plans as static lists, and control token usage and latency despite ever-growing chat history. As a companion product, responses needed to feel immediate and conversational during planning flows, and the experience had to be available not only on the web but through a Chrome extension while users browsed or worked.

Surfaces

Tokens

Core Constraint

Long-Term

Memory

Services Delivered

AI Integration MVP

AI Companion for ADHD Task Planning & Executive Function Support

Architecture Overview

Data Layer

Neon Serverless Postgres Auto-Compacted Memory

Backend & Orchestration

FastAPI LangGraph OpenAI Stripe

Frontend

SvelteKit Chrome Extension

Infrastructure

GCP Docker

Key Technical Decisions

System Design

Comigo was built as a web application with an AI orchestration layer and a browser-based companion experience. A FastAPI backend handled API services, user sessions, task workflows, and AI orchestration, while a Svelte/SvelteKit frontend kept the interface lightweight and responsive. A Chrome extension extended the companion into the browser so users could interact while working. A LangGraph-based agent workflow drove streaming responses, tool calling, and structured multi-step reasoning over OpenAI models, with serverless Postgres on Neon storing users, chat logs, task records, and summarized memory. Stripe powered secure payments and subscription billing.

Key Decisions

SvelteKit was chosen to keep the frontend — and especially the Chrome extension — lightweight, reducing bundle weight and improving responsiveness versus heavier client-side frameworks. LangGraph was selected for agent orchestration to support streaming and tool calling, moving the AI beyond simple chat completion into controlled execution paths. Rather than passing full chat history into every request, the system compacted historical conversations into milestones, preserving useful context while cutting token usage and latency. Deployment was intentionally kept simple: the application was containerized with Docker and run on Google Cloud Platform, while a serverless Postgres database on Neon autoscaled with load and removed manual provisioning — reducing operational overhead so the team could prioritize validating the product experience over infrastructure.

Implementation Highlights

Streaming AI responses made the companion feel immediate and interactive. Tool calling through LangGraph supported structured workflows beyond plain conversation. Chat history was persisted to maintain continuity across sessions, while auto-compacted memory summaries reduced token usage without losing relevant user context. Dynamic task reprioritization responded to changing user input, task-planning flows sequenced work into smaller manageable steps, and Chrome extension integration kept the assistant accessible during active browsing. Secure payment options were integrated via Stripe, adding subscription billing while keeping sensitive card data off Comigo's own servers.

Results & Validation

Reached MVP and launched publicly on Product Hunt and AppSumo within roughly two months.

Delivered a working AI companion experience for ADHD-oriented task support, with persistent chat history and memory-aware interaction.

Validated task planning and reprioritization driven by evolving user context rather than static lists.

Shipped streaming AI responses with tool-calling workflows for a real-time, conversational feel.

Made the assistant accessible across both web and a Chrome extension, with auto-compaction of conversation history for token efficiency.

Key Insights

An ADHD-focused assistant cannot behave like a static task manager — usefulness depends on understanding context, preserving continuity, and adapting to changing priorities without overwhelming the user.

With a memory-aware assistant, the hard problem is not storing history but making it compact, relevant, and cost-efficient when fed back to the model.

Agent workflows that combine chat, task planning, and tool execution require tight product-engineering alignment — output quality depends on when, how, and what context is passed into the model.

Memory compaction is an effective lever for managing long-context constraints, retaining continuity while controlling token cost and latency.

A lightweight stack — SvelteKit plus a focused LangGraph agent layer — can deliver a consistent experience across both web and browser-extension surfaces.

Who This Applies To

This architecture is relevant to companies building AI companions, productivity agents, mental-wellness tools, coaching platforms, or workflow assistants that need persistent memory and adaptive task execution. The same approach applies wherever users need ongoing, context-aware support rather than one-off chatbot responses.

AI Companions Productivity Agents Mental Wellness Memory-Aware AI Conversational UX

Technologies Used

Backend

FastAPI LangGraph

Frontend

SvelteKit

Infrastructure

GCP Docker Neon Serverless Postgres

Data & Integrations

OpenAI Chrome Extension Stripe

Patterns & Techniques

Streaming Responses Tool Calling Auto-Compaction Long-Term Memory Dynamic Reprioritization

Tools

GitHub

Building something similar?

We specialize in ai integration and mvp for productivity / wellness companies. If you're facing challenges like the ones we solved for Comigo, let's talk.

Book a Free Consultation → View All Case Studies

30 minutes · No commitment · We respond within 24h