Case studies & technology

Real outcomes from private AI

Client deployments that eliminated token billing, passed compliance review, and shipped production AI agents — plus the open-source stack we build on.

What we've shipped

Outcomes, not slide decks. Every engagement runs on the same private, local-first stack we research and use ourselves.

$2,500/mo → $0

Insurance SaaS replaces closed-API generalist with a 14B specialized model

Beat GPT-4o on the task. 4× usage, zero increase in spend.

Qwen2.5 14B · LoRA · Self-hosted

Read case study

Privacy-first compliance

AI pay-stub discrepancy detection on a self-hosted stack

Gemma 3 27B on Courier OS — sensitive payroll data never leaves client infrastructure.

Gemma 3 27B · Courier OS · OpenAI-Compatible

Read case study

~$100/mo all-in

Cook'n App ships AI recipe features on Courier Cloud

Production AI cost capped at a predictable monthly bill — not per-token roulette.

Courier Cloud · n8n

Read case study

Client work

Production AI that cuts costs and ships

Strategy, implementation, and ongoing optimization for teams migrating off cloud token billing or deploying compliant local agents.

Active

Pharmaceutical AI SaaS

A New Software Platform for AI Powered FDA Feedback

Client is building a new SaaS for pharmaceutical companies that uses AI for FDA-backed feedback to document submissions.

RAGSelf-hostingDevelopment

View project

Active

AI Support Agent

Developing a customer support agent to offload support reps

Client has 2 support reps that are overwhelmed with support requests. We are working with them to create an AI support agent that will live on their website to offload support requests with instant AI answers.

MeeblySelf-hostingWeb

View project

Active

Self-Hosted Automation

Client wanted to automate internal processes to free up time.

We worked with a small but established business owner who was very data concsious and his own bottleneck in his business to creat private, self-hosted automations that free up his time.

n8nCourierMac Mini

View project

Past Project

Insurance AI Automation

A 14B specialized model that outperformed GPT-4o — on hardware the client owns.

Our flagship case study for the specialized-models thesis. The client was paying ~$2,500/mo for AI on a closed-API generalist and wanted to 4x their usage. We LoRA fine-tuned Qwen2.5 14B for the specific task and it outperformed GPT-4o on their workload, eliminated their token-based spending, and gave them headroom to scale without their bill scaling with them.

TensorRT LLMOpenAI-CompatibleQwen2.5 14B

View project

Past Project

AI Pay Stub Compliance Product

A product to help employees find payment discrepancies that leveraged Courier self-hosting.

We worked with a founder developing a payments tracking app that used AI to search for and discover complex payment discrepancies. Due to privacy concerns the AI had to be self-hosted. We used Gemma 3 27B for the task and it performed very well.

vLLMOpenAI-CompatibleGemma 3 27B

View project

Past Project

AI Features in Cook'n App

Features users loved in a beloved recipe organizer going on over 30 years in business

We worked with DVO Enterprises and their Recipe Organizer app Cook'n to create a robust AI strategy for features users would love, and assisted their developers with implementation.

StrategyIntegration SupportCourier Cloud

View project

Courier OS

The operating system for local AI

Courier OS is our own software suite — built to make local AI production-ready. Every engagement runs on it: a web dashboard for management and monitoring, plus inference, agent interfaces, custom apps, automations, agentic coding tools, and device management on dedicated hardware you control.

What we build on Courier OS Web dashboard

Jackson Oaks and Ryker Ross — RecursionAI co-founders and the team that built Courier OS.

Open source

Tools we build in the open

Agent interfaces, MCP servers, and infrastructure primitives from the RecursionAI lab.

Active

FlowDB

AI-Native Database Stack. Built with AI in mind for easy integration.

FlowDB is an open-source, AI-native database stack built for the age of agents. JSON in, JSON out — with automatic vectorization for RAG, MCP compatibility out of the box, and a REST API for programmatic control. Dockerized for painless deployment.

PythonNoSQLVector DB

View project

Active

Courier Dispatch

MCP Server for Guided Coding Assistance. Makes AI coding assistance easy and safe.

Courier Dispatch is an open-source MCP server that turns any AI chat app into a guided coding assistant. It teaches, reviews, and verifies — but never writes code for you. Read-only codebase access, structured task management, and constrained command execution keep the developer in the driver's seat.

PythonMCPCLI

View project

Past Project

Mosaic

AI Built for Families and Teams

Mosaic is a cross-platform AI chat app built with Kotlin Multiplatform, designed for families and teams. All the best AI models in one app, with guardrails, usage limits, and conversation monitoring. Parents can manage their kids' usage, monitor chats, and create rules — helping the next generation master AI safely.

Kotlin MultiplatformTypeScriptMulti-model AI

View project

Past Project

Pseudo Engine

Programming Language for Structured AI Code-Generation

The Pseudo Engine is the first AI transpiler, built in Rust, with a custom programming language called Pseudo Syntax. Write in Pseudo Code — a concise language designed for structured code generation — and the engine transpiles it to any programming language using a fine-tuned AI model. Includes a CLI, version control system, and IDE plugins.

RustCustom LanguageCLI

View project

Estimate your token savings