Sell Your Logic
  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
Level 4 · AI agent with tool calling, memory and human confirmation

FindPropertyDealers · Rehnuma, the property guide

Let the model choose the tools, and verify what it did

A visitor types "5 marla plot near Islamabad airport, up to 25 lakh". The model chooses which of our functions to call (search societies, find dealers, draft a requirement), reads the results and decides the next step. Our code builds every card.

The business need

Buyers want the right society and a trustworthy dealer, in Urdu, Roman Urdu or English. Nothing can be invented: no dealer, price or approval status, and nothing is sent to a dealer without the visitor.

12 / 12

evaluation cases passed before release

1.7 s

median answer time in that run

5

failures found by testing the real model, each fixed in code

How one request flows

From the visitor's words to what they see, step by step.

  1. 1

    Guard

    Personal details never reach the model.

  2. 2

    Memory

    The conversation so far, kept for an hour, with a turn limit.

  3. 3

    Model chooses a tool

    For example: search societies in a city, or find verified dealers.

  4. 4

    Our code runs it

    A typed function over our database. The model never touches the database.

  5. 5

    Loop

    The model reads the result and calls another tool or answers, within a step limit.

  6. 6

    Cards and confirm

    Code renders the cards. A draft is sent only when the visitor confirms.

The key ideas

Tools are the only way to know

The model can only learn about societies and dealers by calling our functions. Each has a name, a description and typed parameters.

Cards come from code

Dealer and society cards are built from tool results, never from the model's words, so a name, price or approval cannot be invented.

Write actions need a human

"Draft a requirement" only prepares a draft. The visitor reviews it and presses a button before anything reaches a dealer.

Data is not instructions

Dealer bios and visitor text arrive as data. Even a bio that says "ignore your rules" cannot change what the cards show.

Verify, don't trust

If the model says "your draft is ready" without calling the tool, code notices, asks once more with only that tool, and if needed builds the draft itself.

Slot filling with rules

City, size, budget and purpose are read from the visitor's own words with exact rules ("25 lakh" = 2,500,000), as a safety net for the model.

Limits on everything

Characters per question, model calls and tool calls per question, requests per minute and a daily cap.

Offline is guided

With no AI, the chat becomes buttons: city, then society, then dealers. A visitor is never stuck.

What testing taught us

Found by running the real model, not by guessing. This is the most useful part.

It asked permission instead of acting

The model replied "Shall I search?". A prompt rule fixed it: never ask permission to search, just search.

It said it did something it did not do

It wrote "your draft is ready" without calling the tool. The visitor would have got nothing. Code now verifies after the loop and repairs it.

It guessed, confidently and wrong

It assumed a society was in Lahore and listed the wrong city's dealers. The search tool was redesigned to work without a city and to say where a name was found.

It copied the cards into text, in the wrong language

Code now cuts pretend cards out of the reply and detects the language of the latest message.

The fix is code, not a longer prompt

A small model fails in these predictable ways. Each fix above lives in code, where it can be tested.

What you learn at this level

  • Agents and tool schemas
  • Multi-step reasoning
  • Conversation memory
  • Human confirmation for actions
  • Prompt-injection defence
  • Tracing what an agent did

Where it stops

The agent talks directly to one model family. The next level runs the same pattern on any provider.

Open now · free

AI Playground

Change chunk size, similarity threshold, model and prompt on our case studies, and see what was retrieved, the exact prompt, the answer, the speed and the cost.

Open the Playground →
Live · paid with tokens

Live workshops

Hands-on sessions that build each level step by step, hosted by our founder and by experts on SellYourLogic.

See the workshops