Agents that do the work, not just answer questions.
iLeaf builds production AI agents that plan a task, call your real systems and finish it. We have shipped multi-agent orchestration, Graph-RAG retrieval and voice agents into regulated environments since 2011 — each one running under AgentOps, where policy, cost and safety are enforced at runtime rather than promised in a document.
Ask Lia about ai & agentsWhat this includes
- Multi-agent orchestration
- Sequential and swarm agents that divide a task, call tools and hand off to each other, with a supervisor that can stop the run.
- Graph-RAG retrieval
- Retrieval over a knowledge graph rather than flat chunks, so the agent can follow relationships instead of guessing from nearby text.
- Tool and API integration
- Your existing endpoints wrapped as MCP skills an agent can call safely, with scopes and rate limits per tool.
- Voice agents
- Autonomous phone handling on real numbers — qualification, follow-up and escalation with barge-in support.
- Evaluation harnesses
- Task-level test suites so a prompt or model change is measured against your cases before it reaches production.
- Governor Engine
- Runtime limits on spend, latency and permitted actions, with escalation to a human instead of a confident guess.
How we run it
Every phase ships something usable on its own, so you are never holding a half-finished system waiting on the next milestone.
Prove the use case is worth automating
We size the task, the volume and the cost of being wrong. If conventional engineering solves it more cheaply, we say so — we once saved a client ₹12 lakh by recommending exactly that.
Ground it in your data
Before any agent is written we build retrieval that answers correctly on your own documents and records, and measure it.
Give the agent a narrow first job
One task, read-only where possible, with an approval gate. It earns wider permissions by being right.
Instrument, then widen
Cost, latency, tool errors and escalation rate go on a dashboard from day one. Scope grows only where the numbers hold.
What we build it with
- Model Context Protocol (MCP)
- LangGraph
- Python
- TypeScript
- Graph-RAG
- Neo4j
- pgvector
- Claude
- OpenAI
- Open-weight models
- Temporal
- OpenTelemetry
Questions we get asked
How is an agent different from a chatbot we could buy?
A chatbot returns text. An agent plans a multi-step task, calls your real systems and completes it — creating the ticket, drafting the offer, updating the record. That means it needs permissions, limits and an audit trail, which is most of the engineering work and the part off-the-shelf tools leave to you.
What happens when the agent gets it wrong?
It escalates rather than guesses. The Governor Engine enforces what an agent may do, how much it may spend and when it must stop, and every action is logged with the reasoning behind it. Failures are held and reviewed, never silently retried.
Do you need our data to train a model?
No. We use retrieval over your data rather than training on it, so nothing is absorbed into model weights and nothing is sent to train public models. Where a client requires it, we have run the entire stack on their own on-premise GPUs.
Let’s talk about ai & agents.
Tell us what you are running and what it needs to do next. We will tell you honestly whether we are the right team for it.