Ballast
A high-performance AI context engine that gives your agents real data without the headaches.
About Ballast
Ballast is a high-performance AI retrieval engine built to solve one problem. Getting your AI agents to use your real data without slowing down, losing context, or forcing you to maintain a zoo of services. It runs as a single Go binary, ships with an embedded SvelteKit UI, and includes native MCP tools so any agent can instantly search, sync, and retrieve your company’s knowledge.
The goal is simple. Make retrieval boring and predictable so your AI agents can finally behave like they have a memory instead of amnesia.
My Role
I built Ballast from the ground up. Architecture, code, docs, MCP design, integrations, Temporal workflows, SvelteKit UI, everything. It started as a personal need for a retrieval layer that didn’t break under real workloads, and grew into a full platform that handles sync, search, embeddings, indexing, rate limiting, and resilience automatically.
My responsibility was to make retrieval feel effortless for developers and fast for agents while keeping the entire system simple to deploy and maintain.
Technology Stack
- Go (go-zero framework)
- SvelteKit + TypeScript
- Qdrant (vector search)
- PostgreSQL
- Redis
- Temporal
- NATS
- OpenAI, Cohere, Jina (embeddings and reranking)
Key Achievements
- Built a single-binary retrieval system with real-time sync and 42 data source integrations.
- Designed and implemented a native MCP server so Claude and other agents can use Ballast with zero setup.
- Created a hybrid semantic search engine with fusion, reranking, smart filters, and sub-10ms latency.
- Shipped a clean, embedded Web UI for management and debugging without adding any extra services.
Links & Resources
Related Projects & Content
Introducing GoBot, an Autonomous AI Agent That Actually Controls Your Desktop
GoBot is a persistent AI agent that remembers conversations, controls your desktop, runs scheduled tasks, and works across multiple channels. Tonight I watched it check my Gmail, compose an email, and click Send. All by itself.
I Reduced My MCP Tools from 96 to 10. Here's the Pattern.
MCP tool bloat is choking LLMs. The Single Tool Resource Action Pattern (STRAP) applies REST-style routing to consolidate 96 operations into 10 tools with the same functionality.
How to Use Ballast to Bypass the Usual Headaches of Getting Your AI Agents to Use Real Data
I've hit my limit with AI systems that almost work. In demos everything looks amazing, then the moment you point an agent at real company data, everything slows down, breaks, or vanishes. Let's fix that today with Ballast.