Back to portfolio

Octaria

Octaria

About Octaria

Octaria is an AI-native software agency based in Houston, Texas that has delivered custom software to small and mid-market companies since 2018. The firm builds with agents, runs on agents, and focuses on three areas: applied AI agents for business automation, internal AI operations (model selection and governance), and AI infrastructure cost management. I work with Octaria as a contract Software Applications Engineer, embedded on a long-running engagement for one of their clients—an energy investment firm.

About The Project

Mission: Build and operate a production AI agent platform for an energy investment firm—giving departments, crews, teams, and individual employees their own AI agents to help with everyday operations, research, and reporting. Because the firm works with highly sensitive deal and financial data, the platform had to make agents genuinely useful while guaranteeing at the network level that private data can never leave the company.

Technologies Used: Anthropic Claude (Opus, Sonnet, Haiku), OpenClaw (Agent Runtime), Model Context Protocol (MCP) Servers, Next.js, React, TypeScript, Tailwind CSS, Shadcn, TanStack Query, Node.js, Python, AWS (EC2, S3, Systems Manager, Secrets Manager, DynamoDB, CloudWatch, WAF, Route 53), AWS CDK, Terraform, GitHub Actions (CI/CD), Microsoft Teams (Bot Framework), Microsoft Graph, Microsoft Entra ID, Slack, Squid (Egress Proxy), Tailscale (VPN), LanceDB (Vector Memory), Docker, Vitest, MkDocs Material (Technical Documentation), Claude Code (AI-Assisted Development with Custom Skills & Commands), DocDD (Document-Driven Development).

Links

Summary

On this project I serve as the primary engineer on the platform, contributing across the agent runtime, the admin dashboard, cloud infrastructure, messaging integrations, and documentation. The work spans everything from provisioning and securing individual agent machines to the tooling that lets a small team ship and operate a growing fleet of agents safely.

Closed Zone / Open Zone Security Model: Working with the project's network architect, I implemented a two-zone architecture on AWS. “Closed Zone” agents can work with private company data and the CRM but have no path to the public internet—outbound traffic is locked down at the network layer with security groups and an allowlist egress proxy, so that even a fully compromised agent cannot exfiltrate data. “Open Zone” agents are the opposite: full internet access for public research, and no access to private data. Data flows one way only, from Open to Closed. The agent runtime is deliberately treated as untrusted; the network is the guard rail, and application-level rules are a second line of defense.

Agent Provisioning Dashboard: Shipped a Next.js and TypeScript admin dashboard that provisions agents end-to-end—launching the machine, delivering a signed configuration bundle, bootstrapping the agent runtime, registering it with messaging channels, and reporting live progress back to the UI. What had been a 2–4 hour manual checklist per agent now takes roughly 90 seconds and requires no operator shell access. The dashboard integrates with Microsoft Entra ID so an administrator can select any employee and provision their personal agent exactly once.

Agents Installed Anywhere: Built the gateway layer that lets agents live wherever employees already work—Microsoft Teams (a multi-bot relay supporting direct messages, channels, file upload and download via Microsoft Graph, and voice), Slack, and web chat. New agents are registered with the relay automatically at provision time, and each deployment is verified against a QA matrix covering files, permissions, and voice on every surface.

Recurring Research Workflows: Employees and departments can set up scheduled agent jobs that gather news and market data from a variety of APIs and sources, then deliver recurring reports that help them track the industries they are making investment deals in. Data sources are exposed to agents through custom Model Context Protocol (MCP) servers written in Python and Node.js.

Agent Memory & Shared Knowledge: Implemented layered agent memory—individual memory on each agent's machine, plus a shared vector store (LanceDB) that lets knowledge learned by one agent benefit the rest of the fleet, with a curation step that gates what is promoted into the shared layer.

Documentation-Driven “Software Factory”: Built a documentation engine that Claude Code both references and keeps current as it works: 40+ Architecture Decision Records, 30+ technical specifications, product requirements, runbooks, QA reports, and progress reports, all served through MkDocs. Custom Claude Code skills encode the team's workflows—writing ADRs, reading and writing documentation, security reviews of anything that touches networking, and finishing pull requests—so AI coding agents work from the same source of truth as the humans.

Infrastructure & Operations: Own the infrastructure as code (AWS CDK and Terraform), GitHub Actions deployment pipelines for the dashboard and documentation sites, fleet health monitoring with agent heartbeats and alerting, backups, and version-controlled agent workspaces so that git remains the durable record of what every agent is configured to do.