AI agents and automation
AI agents and automation that run on their own
Most AI prototypes stall when they meet real traffic, rate limits and messy data. I build the version that runs every day: workflows on job queues with retries, agents that read and act on your live systems, and a person in the loop where a mistake would cost money.
Is this for you
A good fit if
- Your team repeats the same research, replies or data entry every day and you want it done automatically.
- You have an AI prototype that works in a demo but breaks on real volume.
- You want an assistant that answers from your own documents or database, not from guesses.
What I build
What this looks like in practice
Workflow automation
LLM steps moved out of request handlers and into queues (BullMQ or Postgres), with retries that respect provider rate limits and output validated against a schema.
Agents that act on live data
Agents that look up orders, stock or accounts through real APIs, take actions, and hand over to a person when they're unsure.
Answers from your own content
Retrieval over your documents with pgvector or a vector database, citations back to the source, and evaluation so you know how often it's right.
Review steps and cost tracking
Approve-before-send queues for anything customer-facing, and per-run cost so you can see what the automation spends.
How it works
From first call to handover
A short call
You describe the problem, the systems you already have and the deadline. I ask the questions that decide the shape of the build.
A written scope
You get a plain document with what I'll build, what I won't, the milestones and the estimate, before any work starts.
Weekly progress
I build in small, working increments you can try, with a short update each week. Nothing disappears for a month.
Handover
Code in your repository, deployed on infrastructure you control, with documentation your team can run it from.
What you get
At the end
- A working automation or agent in production, not a notebook
- Monitoring for failures and spend
- A runbook for your team: how it works, how to change prompts and limits
- The code in your repository
Stack
Usually built with
- OpenAI and Anthropic APIs
- AI SDK
- Node.js
- Python
- BullMQ
- PostgreSQL
- pgvector
Related work
Things I've built along these lines
SaaSBanowi AI Commerce AgentBanowi is an AI sales and support agent for Shopify merchants on WhatsApp. It answers customers from live catalogue data, places and tracks orders in the chat, matches payment screenshots to orders and hands off to a human in one click.
SaaSReelspanReelspan is a social publishing API and dashboard that posts one video, photo set or text post to YouTube, TikTok, Instagram, X, LinkedIn and more at once.
Open SourceForgekitForgekit is an open-source Next.js 16 AI starter kit: a streaming AI chatbot with tool calling, five model providers including local Ollama, RAG on pgvector, an MCP endpoint, Better Auth and a Docker self-hosting path.
How I approach it
Guides I've written on this
- Building an AI Automation Pipeline in Node.js with BullMQ: Queues, Retries and Rate Limits6 min read
- How to Build RAG with pgvector and the AI SDK in TypeScript5 min read
- pgvector vs Pinecone vs Qdrant: Benchmarked for RAG in 20269 min read
- Agent Skills Explained: How to Write a SKILL.md for Claude Code, Cursor, Codex and Copilot11 min read
Tell me what you need built
Send the problem and a rough timeline. I reply personally, usually within a day, with questions or a next step.