What is an AI agent?
An AI agent is a program that uses a language model to understand a request, decide which tools or data it needs, and return a result. A chatbot only talks. An agent also looks things up, runs queries, reads documents and, where you allow it, takes an action.
Restrologic is an Indore engineering studio. We build agents that sit on top of real business data and run every day. We do not sell a generic chatbot with your logo on it.
What we build
We build three kinds of agent most often.
Analytics assistants come first. Owners and managers ask questions such as “which outlets sold less this week than last week, and why?” and get an answer with a chart. Behind the answer, the agent writes SQL, and our checker parses that SQL before it runs. Only read queries are allowed, and only on the client’s own stores. The assistant also searches documents you upload, so a question can draw on a sales table and a contract in one answer. Our own version of this is Reckoner, which has been live since 30 August 2026.
Document intelligence is the second. Many businesses hold years of reports, decks and scanned PDFs that nobody can search properly. We build a knowledge graph from those files in which facts carry dates. The agent knows that a person was Associate Director in 2023 and Vice President from September 2026, and it answers for the date you ask about. It reads scanned pages and charts with a vision model, merges name variants such as a short and a full spelling of the same person, and cites the file and page for every answer. Most questions cost about $0.001. The full story is in our document intelligence case study.
Screening is the third. For complaint handling we built a screening step that checks photos for signs of reuse or editing. It does not decide anything. A human approves every refund. The agent shortens the queue, and a person keeps the final call.
How does an agent work, step by step?
A good agent is a small set of controlled parts. This is the shape we use.
- A question arrives in plain language, from a web console, a phone app or a messaging channel.
- The agent picks a tool: a database query, a document search, or a code sandbox.
- The tool runs inside limits. SQL is parsed and must be read-only. Code runs in a throwaway sandbox with no network access.
- The agent turns the result into an answer, with a chart or a citation.
- Corrections from users are stored, and an AI judge scores the answer so quality is measured over time instead of guessed.
We also build and use MCP tool servers. MCP is a standard way to give an agent a defined list of tools. It lets us add a new capability, such as a lookup against your ERP, without rewriting the agent.
What proof do you have?
| Agent | Status | What it does |
|---|---|---|
| Analytics assistant | Live since 30 August 2026 | Plain-English questions, answers with charts, read-only SQL limited to the client’s stores |
| Document intelligence | In production for an unnamed client | Dated facts, scanned PDFs, answers cited to file and page |
| Complaint screening | Built | Checks complaint photos for reuse or editing, a human approves every refund |
| Vedikshana | Live consumer app | Vedic astrology guidance in English, Hindi and Hinglish |
Vedikshana is our test of safety at consumer scale. It has a 147-case safety test suite, crisis detection that changes how the app responds, and a cost cap of 3 cents per message. Those numbers come from the code, and we are glad to walk a buyer through them.
Who is this for?
AI agents suit businesses that already hold useful data and lose time getting answers from it.
- Restaurant and retail groups with sales data spread across outlets and channels.
- Agencies, law firms, finance teams and research groups with large document archives.
- Product companies that want an assistant inside their own app.
- Operations teams who spend hours on routine screening before a person decides.
If you have no data yet, an agent is the wrong first step. We will say so, and suggest custom software to capture the data first.
How does an engagement run?
Most projects follow the same four stages.
- Question list. We collect the questions your team asks most often, and the answers you would trust. This becomes the test set.
- Data access. We agree what the agent may read, what it may never touch, and where the data lives. For client-facing agents we limit access per tenant.
- Build and test. We build the agent, run it against the test set and show you where it fails. You see real answers early, not a slide deck.
- Launch and tuning. We go live with a small group, collect corrections, and tune the prompts, tools and model choice. Support continues after launch.
You choose the model. We can run with Claude, GPT, DeepSeek or Qwen through OpenRouter, and we can change it later as prices and quality move.
How we keep agents safe
Safety is a build decision, not a policy document. Our agents follow these rules.
- Read-only by default. Writes happen only through a named tool with its own checks.
- Tenant limits. An agent for one client cannot see another client’s rows.
- No network from code sandboxes. Code that the agent writes cannot call out.
- Human approval for money, such as refunds.
- Cited answers over documents, so a reviewer can verify each claim.
Next step
Send us the three questions you wish you could ask your data today. Talk to Deepak and we will tell you whether an agent is the right fix, what it would read, and what a first version would look like.