Services / AI systems
Custom AI Agent Development
AI workflows need to be useful, observable and safe to operate. I build the surrounding product and infrastructure that make model-driven features dependable.
What I can build
I develop AI features inside real products: agents that call tools, coordinate multi-step work, integrate with external APIs and report progress to users. A useful agent needs boundaries, observability and a recovery path, not just a prompt.
- LLM and provider integration for an existing web or SaaS product
- Structured tool calling and multi-step workflows
- Job queues for slow or expensive generation tasks
- Human review, validation and approval points
- Streaming status and retry handling for long-running work
- Usage tracking, caching and cost controls
Architecture and safeguards
For Reevera, model output proposes commands while the application validates those commands against schemas, budgets and locks before changing PostgreSQL state. BullMQ and Redis isolate media jobs from the API, and Server-Sent Events report progress. This pattern keeps probabilistic model output outside the trusted data layer.
The right design depends on the use case. A knowledge assistant may need retrieval and source checks; a workflow agent may need strict tool permissions and explicit human handoff. I scope those decisions with the client rather than promising a generic agent will fit every process.
Technology in published work
My Reevera work uses TypeScript, Fastify, PostgreSQL, Redis, BullMQ, FFmpeg and provider adapters for Gemini, Imagen, Veo, ElevenLabs, OpenRouter and Claude. MetaLLM work includes model routing and debate workflows. The specific provider for a new project is selected around its requirements, cost and data needs.
Evidence
The Reevera case study details queue isolation, AI command validation and real-time state recovery. The MetaLLM case study covers model routing, debate workflows and the SaaS product around them.
How a project starts
Share the workflow you want to automate, the tools and data it must access, expected volume, error tolerance and who approves consequential actions. I can use that to define an initial, testable scope.
Have a project in mind?
Let’s talk about an AI workflow.
Share your goals, current system and timeline. I can help scope the architecture and implementation.