← Back to Portfolio
AI Agents & Tooling
AI Agent Platform: A Virtual Workforce
Enterprise client — under NDA10 agents · 21 skills · 205 tool methods
AI / LLM AgentsAgent ToolingNext.jsReactTypeScriptPostgreSQLDrizzle ORMKubernetesAWS
An enterprise platform that gives a company a virtual workforce — a fleet of AI agents that do real business work around the clock, while the humans stay at the edge doing the judgment, relationship, and creative work that actually needs a person.
The premise is that most of what a growing company spends on isn't the work itself, it's the coordination around the work: status meetings, email chains, manual handoffs, someone re-typing a number from one system into another. That layer is what the agents replace.
The agents. Each one is specialized by function rather than being one general-purpose chatbot — revenue and pipeline, operations, marketing, finance, customer success, executive assistant, communications, and a QA agent whose job is watching the other agents' output for drift. They have persistent memory, they share context with each other, and they run continuously instead of waiting to be asked. People talk to them in chat, watch what they're doing, and set what they're allowed to touch.
Organizations and sub-organizations. The whole platform is scoped to an org hierarchy, and authority inherits downward. A parent org's agent can act across its children; a child org's agent can't see sideways or upward. That scoping isn't a UI convention — it's injected at the data layer, so an agent can't ask for another organization's data even if it tries. For a product where the whole value proposition is handing autonomous software real access to your business, that boundary is the thing that makes it adoptable.
Tools and skills — this is the part I build. The distinction matters: tools are the APIs, the concrete things an agent is capable of doing, and skills are how and when to use them — the instructions that turn raw capability into judgment. An agent with a tool can technically call it; an agent with the matching skill knows it should. The platform runs 37 tool groups across 205 methods, with 21 skills teaching 10 agents how to wield them.
Those tools come in three flavors:
Internal — the company's own operating surface: tasks, tickets, deals and pipeline, metrics, agent management, and the dashboard widgets people read.
Integrations — the systems a business already runs on: Google Workspace, Salesforce, Intuit, calendar and email.
Creative — the ones that make an agent able to actually produce something rather than just report: HeyGen for video, ElevenLabs for voice, Brandfetch for brand assets.
Every one of them is parsed and validated at the boundary, so a malformed agent call fails loudly instead of quietly writing garbage into a live business system. There's a test asserting that every tool method has a skill granting it — closing the gap between what an agent can reach and what it's been taught to reach.
One thing worth calling out. When a language model narrates its own tool results back to you in prose, it miscounts rows and invents entries — and a virtual workforce nobody can trust the numbers from is just an expensive demo. So the product renders the structured payload the tool actually returned, one-to-one, as real components. Counts are a literal array length. The model never touches the data. That took a persistent class of data errors to zero and cut the token cost at the same time, since the model no longer re-types result sets it already fetched.
Built on Next.js and React with a TypeScript, PostgreSQL, and Drizzle data layer, deployed on Kubernetes and AWS.
The premise is that most of what a growing company spends on isn't the work itself, it's the coordination around the work: status meetings, email chains, manual handoffs, someone re-typing a number from one system into another. That layer is what the agents replace.
The agents. Each one is specialized by function rather than being one general-purpose chatbot — revenue and pipeline, operations, marketing, finance, customer success, executive assistant, communications, and a QA agent whose job is watching the other agents' output for drift. They have persistent memory, they share context with each other, and they run continuously instead of waiting to be asked. People talk to them in chat, watch what they're doing, and set what they're allowed to touch.
Organizations and sub-organizations. The whole platform is scoped to an org hierarchy, and authority inherits downward. A parent org's agent can act across its children; a child org's agent can't see sideways or upward. That scoping isn't a UI convention — it's injected at the data layer, so an agent can't ask for another organization's data even if it tries. For a product where the whole value proposition is handing autonomous software real access to your business, that boundary is the thing that makes it adoptable.
Tools and skills — this is the part I build. The distinction matters: tools are the APIs, the concrete things an agent is capable of doing, and skills are how and when to use them — the instructions that turn raw capability into judgment. An agent with a tool can technically call it; an agent with the matching skill knows it should. The platform runs 37 tool groups across 205 methods, with 21 skills teaching 10 agents how to wield them.
Those tools come in three flavors:
Internal — the company's own operating surface: tasks, tickets, deals and pipeline, metrics, agent management, and the dashboard widgets people read.
Integrations — the systems a business already runs on: Google Workspace, Salesforce, Intuit, calendar and email.
Creative — the ones that make an agent able to actually produce something rather than just report: HeyGen for video, ElevenLabs for voice, Brandfetch for brand assets.
Every one of them is parsed and validated at the boundary, so a malformed agent call fails loudly instead of quietly writing garbage into a live business system. There's a test asserting that every tool method has a skill granting it — closing the gap between what an agent can reach and what it's been taught to reach.
One thing worth calling out. When a language model narrates its own tool results back to you in prose, it miscounts rows and invents entries — and a virtual workforce nobody can trust the numbers from is just an expensive demo. So the product renders the structured payload the tool actually returned, one-to-one, as real components. Counts are a literal array length. The model never touches the data. That took a persistent class of data errors to zero and cut the token cost at the same time, since the model no longer re-types result sets it already fetched.
Built on Next.js and React with a TypeScript, PostgreSQL, and Drizzle data layer, deployed on Kubernetes and AWS.
Architecture
