The platform in six points
An AI agent platform for business automation
ConnectPilot handles document generation, knowledge retrieval, workflow orchestration, and decision support across your operations.
Every agent is a self-contained module
Each one carries its own system prompt, tools, knowledge scope, and model selection. New agents ship in hours, not weeks.
Model-agnostic by design
We typically build on Anthropic Claude, but the platform supports all major chat models: OpenAI GPT, Google Gemini, Meta Llama, Mistral, DeepSeek, and others.
Retrieval scoped per agent
Retrieval-augmented generation runs on PostgreSQL with pgvector, scoped per-agent so knowledge bases never bleed across agents.
A dedicated deployment for every client
Their own repository, hosting, database, and credentials. No shared infrastructure, no shared data, no shared blast radius.
Deployed where it fits
We deploy on Google Cloud or Render, whichever fits the client's environment.
How AI agents work: what we mean by an AI agent
An AI agent, in our definition, is an LLM-driven program that:
Receives a goal
Receives a goal as natural language plus structured context.
Pulls grounding
Pulls grounding from a retrieval layer: vector search, structured data, external APIs.
Plans and executes
Plans and executes one or more steps using tools (function calls, document generation, sub-agent delegation).
Streams an auditable response
Streams an auditable response back to the user with source citations.
The ConnectPilot agent framework
We built and own our agent framework end-to-end. It ships as a production-ready starter that we customize per client. We chose to own the framework because production AI agents need three things that generic frameworks fight against:
Token-level streaming with backpressure all the way to the browser.
Predictable latency under autoscaling compute.
Auditability: every decision the agent made, traceable in the database.
The agent contract
Every agent conforms to a single small contract:
Adding a new specialist agent means writing the prompt and (optionally) the tools. The registry discovers it automatically. No frontend deploy required to ship a new agent.
The agentic loop
Assemble messages: system prompt + retrieved knowledge context + conversation history + new user turn.
Stream the model's response.
If the model requests a tool, dispatch it (which may call another agent), append the result, and continue.
Stop when the model signals end-of-turn or hits the turn ceiling.
Every decision the agent made is logged with its inputs and outputs.
AI agent architecture: how the pieces fit together
Requests flow from your interface to a single agent service that orchestrates discovery, reasoning, and memory. Every layer is observable and independently scalable.
ConnectPilot Agent Service
Retrieval-augmented generation (RAG) on PostgreSQL + pgvector
Agents are grounded in uploaded knowledge: documents, contracts, templates, reference materials, structured notes.
Ingestion pipeline
Query path
Top-K retrieval via cosine similarity.
Optional hybrid search combining vector similarity with PostgreSQL full-text search, weighted by title, content, and tags.
Retrieved chunks injected into the agent's system prompt with provenance (document title, page, relevance score) so the model can cite sources.
Every assistant message persists its retrieval sources for audit.
Why pgvector
One database to back up, monitor, and restore. Transactional consistency between retrieval and the rest of application state. No additional vendor on the critical path. pgvector handles tens of millions of chunks comfortably, well past what most workloads need.
Dedicated deployment per client
ConnectPilot is not a multi-tenant SaaS. Every client gets a dedicated deployment: their own repository, their own database, their own hosting, their own credentials, their own LLM provider keys. This is a deliberate design choice.
The starter framework is mature enough that a new client deployment goes from zero to production in days, not weeks.
Infrastructure and observability
We deploy on Google Cloud or Render, whichever fits the client's environment. Both give us fast deploys, managed PostgreSQL, and clean autoscaling without the operational weight of running our own Kubernetes.
Model support
ConnectPilot is model-agnostic. We typically deploy on Anthropic Claude because it leads on tool use, long-context reasoning, and instruction following for agent workloads, but the platform supports every major chat model. Clients can mix providers across agents, or standardize on one.
| Provider | Models we support |
|---|---|
| Anthropictypical default | Claude Haiku 4.5, Claude Sonnet 4.5, Claude Sonnet 4.6, Claude Opus 4.6, Claude Opus 4.7 |
| OpenAI | GPT-5, GPT-4o, GPT-4.1, o3, o4-mini |
| Gemini 2.5 Pro, Gemini 2.5 Flash | |
| Meta | Llama 3.3, Llama 4 |
| Mistral | Mistral Large, Mistral Small |
| DeepSeek | DeepSeek V3, DeepSeek R1 |
| xAI | Grok 4 |
Agent authors set a default; users can override per conversation from a server-enforced allowlist. We don't auto-route between models: the author picks a default; the user can step up or down. Auto-routing makes the question why did this response change impossible to answer.
Security
Security is enforced at every layer, from authentication and service-to-service trust to isolation, secret management, and the full data lifecycle.
What makes this platform work
Three properties matter more than any single technology choice.
Frequently asked questions
Let's Build Something Together
One conversation is all it takes to discover where AI can transform your business.
