Skip to content

Latest commit

 

History

22 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

PragatiSetu

Intelligent Data Capture & Schedule-Linking Layer for Infrastructure Project Management

Smart India Hackathon 2026 — SIH26122

An AI-powered system that bridges the gap between planned project schedules and actual site execution by automatically converting unstructured field updates into structured project activities and intelligently mapping them to the correct L5/L6 schedule activities.


📌 Problem Statement

Large infrastructure projects are typically planned and monitored using structured project-management tools such as Primavera P6 and Microsoft Project.

However, actual site progress is often reported through:

  • Daily Progress Reports (DPRs)
  • Site diaries
  • Excel spreadsheets
  • PDFs
  • Free-text updates
  • Supervisor observations
  • Voice-based updates

The language and level of detail used in these reports often differ significantly from the terminology and structure used in the official project schedule.

Example

Planned Schedule

Activity ID: PIP-204-017
Activity: Erect 24-inch Pipeline Section A
Discipline: Piping
Planned Start: 12-Aug-2026
Planned Finish: 17-Aug-2026

Site Report

"24 inch line spool erection completed today."

The site report does not explicitly mention the Activity ID.

The challenge is to automatically understand the field update and determine which planned L5/L6 activity it corresponds to.


💡 Our Solution

We propose an AI-powered Planning-to-Execution Bridge that transforms unstructured site information into structured, schedule-linked project updates.

                    ┌──────────────────────┐
                    │   Project Schedule   │
                    │ Primavera / MS       │
                    │ Project / Excel      │
                    └──────────┬───────────┘
                               │
                               ▼
                     ┌─────────────────┐
                     │ Schedule Parser │
                     └────────┬────────┘
                              │
                              ▼
                    L5/L6 Activity Database
                              │
                              │
                              │
┌─────────────────┐           │
│   Site Reports  │           │
│ PDF / Excel     │           │
│ Text / Voice    │           │
└────────┬────────┘           │
         │                    │
         ▼                    │
┌─────────────────┐           │
│ Data Extraction │           │
│      LLM        │           │
└────────┬────────┘           │
         │                    │
         ▼                    │
┌─────────────────────────────┐
│ Structured Execution Event  │
└──────────────┬──────────────┘
               │
               ▼
       Semantic Retrieval
          + Embeddings
               │
               ▼
      Top Candidate Activities
               │
               ▼
       Validation & Reranking
               │
               ▼
        Confidence Scoring
          ┌────┴────┐
          │         │
       High       Low
          │         │
          ▼         ▼
     Auto Update  Human Review
          │         │
          └────┬────┘
               ▼
        Updated Project View
               │
               ▼
       Gantt / Progress / Alerts

🎯 Objectives

The system aims to:

  1. Automatically extract meaningful execution events from field reports.
  2. Understand construction and infrastructure terminology.
  3. Map field observations to the correct L5/L6 schedule activities.
  4. Handle differences in terminology and granularity.
  5. Assign confidence scores to AI-generated mappings.
  6. Route ambiguous cases to human reviewers.
  7. Maintain an auditable history of every AI-generated update.
  8. Reduce manual effort involved in progress tracking.
  9. Provide near-real-time visibility of project execution against the plan.

🧠 Key Concept: L5 and L6 Activities

Infrastructure projects use hierarchical Work Breakdown Structures (WBS).

A simplified structure may look like:

Project
│
├── Pipeline Construction
│   │
│   ├── Section A
│   │   │
│   │   └── Pipeline Installation
│   │       │
│   │       ├── L5: Install 24-inch Pipeline
│   │       │    │
│   │       │    ├── L6: Spool P-204-07
│   │       │    ├── L6: Spool P-204-08
│   │       │    └── L6: Spool P-204-09
│   │       │
│   │       └── L5: Pipeline Welding
│   │            │
│   │            ├── L6: Joint 101
│   │            ├── L6: Joint 102
│   │            └── L6: Joint 103

The system's primary task is to connect a field observation to the appropriate scheduled activity.


🔍 Example Workflow

Input

A supervisor submits:

"Spool P204-07 erection completed today.
Welding for Section B started."

Step 1 — Information Extraction

The LLM converts the report into structured events:

[
  {
    "equipment": "P204-07",
    "activity": "spool erection",
    "status": "COMPLETED",
    "date": "2026-08-30"
  },
  {
    "section": "B",
    "activity": "welding",
    "status": "STARTED",
    "date": "2026-08-30"
  }
]

Step 2 — Semantic Retrieval

The system searches the schedule for matching activities.

Candidate Activities

1. PIP-204-017
   Spool P204-07 Erection
   Similarity: 0.95

2. PIP-204-021
   Spool P204-08 Erection
   Similarity: 0.71

3. PIP-204-025
   Section B Welding
   Similarity: 0.68

Step 3 — Validation

The system evaluates:

  • Activity description
  • Equipment/asset identifiers
  • Discipline
  • WBS context
  • Date compatibility
  • Location/section
  • Semantic similarity

Step 4 — Confidence Decision

Confidence: 95%
Status: HIGH CONFIDENCE
Action: Auto-update

For an ambiguous case:

Confidence: 61%
Status: LOW CONFIDENCE
Action: Human Review Required

🏗️ System Architecture

1. Data Ingestion Layer

Supports multiple sources:

PDF
Excel
CSV
Text
Site Diary
Voice

The ingestion layer converts these inputs into a common internal representation.


2. Schedule Processing

Project schedules are converted into structured activity records.

Example:

{
  "activity_id": "PIP-204-017",
  "wbs_level": "L6",
  "activity_name": "Erect 24-inch Pipeline Section A",
  "discipline": "Piping",
  "planned_start": "2026-08-12",
  "planned_finish": "2026-08-17"
}

3. AI Information Extraction

An LLM extracts:

  • Activity/event
  • Equipment
  • Location
  • Discipline
  • Status
  • Progress
  • Dates
  • Quantities
  • Remarks

The model produces structured output rather than directly modifying the schedule.


4. Semantic Matching

The extracted event is converted into an embedding.

The system then searches the schedule activity database for semantically similar activities.

Field Update
     │
     ▼
Embedding
     │
     ▼
Vector Search
     │
     ▼
Top-K Activities

This allows the system to handle different wording.

Example

"24 inch line erection completed"

                ≈

"Erect 24-inch Pipeline Section A"

even though the wording is not identical.


5. Context-Aware Validation

Semantic similarity alone is not sufficient.

The system combines:

Semantic Similarity
        +
Equipment ID
        +
Discipline
        +
Location
        +
WBS Context
        +
Date
        +
Schedule Dependencies

to improve matching reliability.


6. Confidence & Human-in-the-Loop

The system never blindly trusts AI output.

High Confidence

Confidence ≥ Threshold

        ↓

Automatic Update

Low Confidence

Confidence < Threshold

        ↓

Human Review Queue

The reviewer can:

  • Approve
  • Reject
  • Select another activity
  • Mark as unmatched

🛡️ Safety & Reliability

A core design principle is:

The AI must never invent a schedule activity.

The system can only:

  1. Match against existing schedule activities.
  2. Mark an event as unmatched.
  3. Request human intervention.

This prevents hallucinated Activity IDs and incorrect schedule modifications.


🔄 Handling Difficult Cases

Similar Activities

Install Pump P-204
Install Pump P-205
Install Pump P-206

If the report only says:

"Pump installation completed."

The system should not guess.

Instead:

⚠️ Ambiguous Match

Multiple possible activities found.

Human Review Required.

Unmatched Activity

If a report contains:

"Temporary access road constructed."

but no corresponding schedule activity exists:

⚠️ No Matching Activity

Event:
Temporary access road construction

Action:
Planner Review

Conflicting Reports

If two reports provide contradictory information:

Supervisor A:
Welding completed.

Supervisor B:
Welding 80% complete.

the system flags:

⚠️ Conflicting Progress Reports

Human verification required.

📊 Proposed Technology Stack

Component Technology
Frontend React / Next.js
Backend Python / FastAPI
LLM Groq API
Embeddings BGE-M3
Vector Search FAISS
Database PostgreSQL
Document Processing Python
OCR PaddleOCR / Tesseract
Speech-to-Text Whisper
Visualization Gantt / Timeline UI
Deployment Docker

The exact model/provider can be changed depending on API availability, latency, cost and deployment requirements.


📁 Dataset Strategy

The prototype dataset will consist of three primary components.

Schedule Dataset

activity_id
wbs_level
activity_name
discipline
location
equipment
planned_start
planned_finish
dependencies

Field Report Dataset

report_id
report_date
supervisor
discipline
raw_text
source

Ground Truth Dataset

report_id
expected_activity_id
event_type
status
confidence_label

🧪 Evaluation

The system will be evaluated using labelled field-report/activity pairs.

Activity Matching

  • Top-1 Accuracy
  • Top-3 Accuracy
  • Precision
  • Recall
  • F1 Score

Information Extraction

Evaluate extraction of:

  • Activity
  • Equipment
  • Status
  • Date
  • Progress
  • Location

Safety Metrics

Measure:

  • False Matches
  • Unmatched Detection
  • Human Review Rate
  • Incorrect Auto-Updates

A key objective is to minimize false activity mappings, because an incorrect schedule update can be more harmful than requesting human review.


📈 Expected Benefits

The proposed system can help project teams:

  • Reduce manual progress-entry effort.
  • Improve schedule-to-execution visibility.
  • Reduce delays between site reporting and schedule updates.
  • Identify unmatched or unexpected work.
  • Detect conflicting progress information.
  • Improve data consistency.
  • Provide an auditable AI-assisted workflow.
  • Enable project managers to focus on exceptions instead of routine data entry.

👥 Human-in-the-Loop Design

The system is designed as an AI assistant, not an autonomous project manager.

             AI
              │
              ▼
      Recommendation
              │
              ▼
      Confidence Score
              │
        ┌─────┴─────┐
        │           │
      Clear      Ambiguous
        │           │
        ▼           ▼
    Auto/Quick    Human
     Approval     Review
        │           │
        └─────┬─────┘
              ▼
       Schedule Update

This provides both automation and human control.


🚀 MVP Scope

The initial prototype focuses on the most important workflow:

Excel Schedule
      +
Text Field Report
      ↓
AI Extraction
      ↓
Semantic Activity Matching
      ↓
Confidence Score
      ↓
Human Approval
      ↓
Updated Gantt / Progress View

Phase 2

  • PDF processing
  • OCR
  • Voice input
  • Multilingual reports
  • Advanced progress estimation
  • Conflict detection
  • Historical analytics

Phase 3

  • Primavera/MS Project integration
  • Enterprise PMIS integration
  • Real-time notifications
  • Project-level analytics
  • Continuous terminology learning

🗺️ Future Scope

The platform can be extended into a broader infrastructure execution intelligence system.

Potential capabilities include:

  • Delay-risk detection
  • Progress forecasting
  • Automated daily progress summaries
  • Contractor performance analytics
  • Resource bottleneck detection
  • Schedule variance analysis
  • Automated management reports
  • Multi-project portfolio monitoring
  • Conversational project queries

Example

"Which activities are behind schedule in Section B?"

The system could respond using the latest validated project data.


🔐 Design Principles

1. AI-Assisted, Human-Controlled

AI recommends; humans retain authority over important schedule changes.

2. Evidence-Based Updates

Every AI-generated update should be traceable to its source report.

3. No Hallucinated Activities

The AI cannot create arbitrary schedule IDs.

4. Confidence-Aware Automation

Only sufficiently reliable matches should be automatically processed.

5. Auditable

Every mapping and modification should have:

Source
Timestamp
Activity ID
AI Decision
Confidence
Reviewer
Final Action

🎯 Project Vision

Bridge the gap between what the project schedule says should happen and what the site reports actually say happened.

Our goal is to transform project monitoring from:

Site Report
     ↓
Manual Reading
     ↓
Manual Mapping
     ↓
Manual Schedule Update
     ↓
Delayed Visibility

into:

Site Report
     ↓
AI Understanding
     ↓
Semantic Schedule Matching
     ↓
Confidence Validation
     ↓
Human Approval
     ↓
Near-Real-Time Project Visibility

🏆 Smart India Hackathon 2026

Detail Information
Problem Statement SIH26122
Theme Smart Automation
Category Software
Organization Oil India Limited

📜 Disclaimer

This repository contains a prototype implementation developed for Smart India Hackathon 2026.

Demonstration datasets may be synthetic and are intended to reproduce the structure and challenges of real infrastructure project-management data.

The prototype should not be considered a replacement for official project-management systems or professional project controls without appropriate validation, security controls and enterprise integration.


⭐ Key Takeaway

SIH26122 is not simply an AI chatbot.

The core innovation is the reliable transformation:

Unstructured Site Information
             ↓
      Structured Events
             ↓
     L5/L6 Schedule Mapping
             ↓
      Confidence Validation
             ↓
       Human Approval
             ↓
       Schedule Update

The primary objective is to create a trustworthy Planning-to-Execution Bridge for large infrastructure projects.


📌 Built for Smart India Hackathon 2026

SIH26122 · Oil India Limited · Smart Automation

About

An intelligent infrastructure management system that converts unstructured field updates into structured progress reports and semantically links them to L5/L6 schedule activities.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Contributors

Languages