MercatoAI

Procurement should start with a conversation, not a form. Our multimodal AI turns voice, photos, and WhatsApp into complete procurement workflows.

Comment

Product Requirements Document (PRD)

ConstructProcure AI

Version: 1.0\ Track: Procurement -- Kaya AI India Hackathon 2026

Vision

Construction procurement should happen before humans realize there is a problem.

ConstructProcure AI is an AI-native procurement assistant that lives alongside existing workflows (WhatsApp, email, ERP, mobile) and automates procurement decisions instead of becoming another dashboard.

It understands documents, conversations, approvals, invoices, and deliveries while proactively preventing delays and payment disputes.

Problem Statement

Construction procurement is fragmented.

Current workflow:

  • Site Engineer raises request
  • Procurement searches vendors
  • Manager approves PO
  • Vendor ships
  • Site receives material
  • Finance performs 3-way matching

Reality:

  • Requests arrive through WhatsApp and calls.
  • Approvals get delayed.
  • Delivery documents are incomplete.
  • Invoice disputes delay supplier payments.
  • Working capital gets locked.

The problem is not a lack of software.

The problem is that information exists across disconnected channels.

Target Users

  1. Site Engineer
  2. Procurement Manager
  3. Project Manager
  4. Finance Team
  5. Vendors

Design Principles

  • Mobile-first
  • AI-first
  • No unnecessary dashboards
  • Human approval only where required
  • Existing workflows over new workflows

User Experience

1. Raise Procurement

User can:

  • Speak
  • Send WhatsApp
  • Upload BOQ
  • Upload RFQ
  • Upload invoice
  • Upload delivery photo

AI extracts:

  • Material
  • Quantity
  • Project
  • Required date
  • Priority

No forms required.

2. AI Understanding

Pipeline:

Voice -> Speech-to-Text

Images -> Vision Models

PDF -> Document Intelligence

Messages -> NLP

Everything becomes structured procurement events.

3. Vendor Recommendation

AI ranks vendors using:

  • Price
  • Historical delivery
  • Location
  • Capacity
  • Payment terms
  • Risk

Outputs:

  • Best vendor
  • Cost
  • ETA
  • Confidence

4. Approval

Instead of sending a raw PO.

Manager sees:

  • Why this vendor?
  • Cost impact
  • Delay risk
  • Alternative options

Approve in one tap.

5. Purchase Order

AI automatically

  • Generates PO
  • Sends email
  • Sends WhatsApp
  • Tracks acknowledgement

6. Delivery Verification

Evidence:

  • Delivery photo
  • GRN
  • Invoice
  • GPS (optional)

AI checks

  • Quantity
  • Damage
  • Missing items
  • Mismatch

7. Finance

Automatic 3-way match

PO

GRN

Invoice

Explains every mismatch.

Core Features

AI Requisition Assistant

Voice + WhatsApp + Documents

Automatically creates PR.

Vendor Intelligence

Vendor ranking

Historical scoring

Risk estimation

AI Approval Copilot

Context-aware approval

Smart PO Generation

Automatic purchase order generation

Delivery Intelligence

Vision verification

OCR

Damage detection

Invoice Intelligence

Field extraction

Mismatch detection

Fraud indicators

Payment Guardian

Tracks payment status

Predicts disputes

AI Architecture

Input Layer

  • WhatsApp
  • Email
  • Voice
  • PDFs
  • Images
  • ERP

↓

Perception Layer

Document AI

Vision AI

Speech AI

↓

Intelligence Layer

Material extraction

Vendor ranking

Risk prediction

Delay prediction

Price intelligence

↓

Reasoning Layer

LLM agent

Policy checking

Recommendation engine

↓

Action Layer

Generate PO

Draft emails

Route approvals

Create reminders

Models

Document Intelligence

Purpose

Understand invoices, BOQs, RFQs, POs, GRNs.

Suggested

  • LayoutLMv3
  • Docling
  • MinerU
  • PaddleOCR

Vision

Purpose

Verify deliveries.

Count materials.

Detect damage.

Suggested

  • YOLO
  • GroundingDINO
  • SAM2

Speech

Purpose

Hands-free procurement.

Suggested

  • Whisper

LLM

Purpose

Reasoning.

Policy understanding.

Vendor communication.

Suggested

  • GPT
  • Llama
  • Qwen

Forecasting

Purpose

Delivery delay prediction.

Suggested

  • LightGBM
  • XGBoost
  • Time-series models

Tech Stack

Frontend

  • Next.js
  • Tailwind
  • React

Backend

  • FastAPI

Database

  • PostgreSQL

Vector DB

  • pgvector

Storage

  • S3 compatible

Queue

  • Redis

LLM

  • OpenAI compatible API

User Flow

  1. User sends request.
  2. AI extracts intent.
  3. AI searches vendors.
  4. AI recommends supplier.
  5. Manager approves.
  6. PO generated.
  7. Vendor notified.
  8. Delivery verified.
  9. Invoice matched.
  10. Payment triggered.

Why This Is Different

Most solutions

  • OCR
  • Chatbot
  • Dashboard

ConstructProcure AI

  • Works where users already communicate.
  • Uses multimodal AI.
  • Automates actions instead of only answering questions.
  • Keeps humans in approval loops.

Future Roadmap

  • ERP integrations
  • WhatsApp business deployment
  • Vendor negotiation agent
  • Predictive inventory
  • Smart glasses integration
  • Autonomous procurement agent

Success Metrics

  • Procurement cycle time
  • Approval turnaround
  • Invoice dispute rate
  • On-time deliveries
  • Manual effort reduction
  • Supplier payment time

Hackathon MVP

Deliver:

  • WhatsApp-like interface
  • Voice requisition
  • Document upload
  • AI vendor recommendation
  • One-click approval
  • PO generation
  • Invoice extraction
  • 3-way match demo

The MVP demonstrates an end-to-end AI procurement workflow without requiring users to adopt a new enterprise process.

Built With

  • agentic
  • ml
  • multimodal
  • osss
  • reasoning
  • vision
Share this project:

Updates