← All services

Service

AI Agent Development

Agents that take real action inside your software

Most AI demos answer questions. The agents I build do work: they look up an order, draft a purchase request, update a record, and know when to hand over to a person.

Each agent is designed around your existing systems and data, with clear limits on what it may do, a log of every step, and people approving anything that matters.

How it works
“Reorder anythingrunning low”AgentorchestratorPlanActReviewERPDocsEmailYou

What I build

  • Multi-agent orchestration

    LangGraph and Semantic Kernel agents that plan, call tools and hand off to each other

  • Customer support agents

    Answers grounded in your own products and records, with humans in the loop

  • RAG over your data

    Search and answers across documents, catalogues and business records

  • ERP-connected agents

    Agents that read and write ERPNext and Frappe workflows safely

What you get

  • A working agent connected to your tools and data, not a demo
  • Tool and API integrations, with clear permissions for each agent
  • Guardrails: drafts instead of irreversible actions, and approval where it counts
  • Conversation and tool-call logs you can review
  • A test set of real requests to check every change against
  • Documentation and a handover session for your team

Watch the agents work

A scripted run on sample data. Pick a request:

  1. Three agents share the work: a Planner breaks the request down, a Builder uses the tools, and a Reviewer checks the result before a person signs it off.

How we’ll work

  1. Discover

    We pick one workflow worth automating, the data it needs and what “done” looks like.

  2. Prototype

    A narrow agent running on your real data, tried with real requests from your team.

  3. Harden

    Guardrails, approvals, logging, cost limits and an evaluation set before anyone relies on it.

  4. Launch and improve

    Release to a small group, read the transcripts, fix what trips it up, then widen.

Omor Al Tanim explaining a workflow on a whiteboard to two teammates
Every agent starts on a whiteboard: the steps, the tools and where a person signs off.

A good fit if…

  • Your team answers the same questions from the same systems every day
  • Work moves between several tools and someone copies data across
  • You have documents, catalogues or records an assistant should answer from
  • You want AI inside your ERP, not one more separate chat app

Tools I use

  • LangChain
  • LangGraph
  • Semantic Kernel
  • LlamaIndex
  • OpenAI
  • Claude
  • Gemini
  • FastAPI
  • Postgres + pgvector
  • n8n
Read: Building multi-agent systems with Semantic Kernel →

Questions

Which AI models do you use?

Whichever fits the job and the budget: OpenAI, Anthropic’s Claude, Google’s Gemini or open models. The agent code stays model-agnostic, so you can switch later.

Is our data safe?

Agents only get the access they need, through your own APIs. Your data stays in your systems, logs live where you choose, and we pick providers and settings that meet your data rules.

What if the agent gets something wrong?

Sometimes it will. That’s why actions that matter become drafts a person approves, every step is logged, and changes are tested against real requests before they ship.

How long does a first version take?

It depends on how many systems the agent touches. A narrow first version is a matter of weeks, not months, and we agree scope and timeline in writing before starting.

Want to build this?

I am available for freelance work and consultations.