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.
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.
Activity ID: PIP-204-017
Activity: Erect 24-inch Pipeline Section A
Discipline: Piping
Planned Start: 12-Aug-2026
Planned Finish: 17-Aug-2026
"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.
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
The system aims to:
- Automatically extract meaningful execution events from field reports.
- Understand construction and infrastructure terminology.
- Map field observations to the correct L5/L6 schedule activities.
- Handle differences in terminology and granularity.
- Assign confidence scores to AI-generated mappings.
- Route ambiguous cases to human reviewers.
- Maintain an auditable history of every AI-generated update.
- Reduce manual effort involved in progress tracking.
- Provide near-real-time visibility of project execution against the plan.
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.
A supervisor submits:
"Spool P204-07 erection completed today.
Welding for Section B started."
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"
}
]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
The system evaluates:
- Activity description
- Equipment/asset identifiers
- Discipline
- WBS context
- Date compatibility
- Location/section
- Semantic similarity
Confidence: 95%
Status: HIGH CONFIDENCE
Action: Auto-update
For an ambiguous case:
Confidence: 61%
Status: LOW CONFIDENCE
Action: Human Review Required
Supports multiple sources:
PDF
Excel
CSV
Text
Site Diary
Voice
The ingestion layer converts these inputs into a common internal representation.
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"
}An LLM extracts:
- Activity/event
- Equipment
- Location
- Discipline
- Status
- Progress
- Dates
- Quantities
- Remarks
The model produces structured output rather than directly modifying the schedule.
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.
"24 inch line erection completed"
≈
"Erect 24-inch Pipeline Section A"
even though the wording is not identical.
Semantic similarity alone is not sufficient.
The system combines:
Semantic Similarity
+
Equipment ID
+
Discipline
+
Location
+
WBS Context
+
Date
+
Schedule Dependencies
to improve matching reliability.
The system never blindly trusts AI output.
Confidence ≥ Threshold
↓
Automatic Update
Confidence < Threshold
↓
Human Review Queue
The reviewer can:
- Approve
- Reject
- Select another activity
- Mark as unmatched
A core design principle is:
The AI must never invent a schedule activity.
The system can only:
- Match against existing schedule activities.
- Mark an event as unmatched.
- Request human intervention.
This prevents hallucinated Activity IDs and incorrect schedule modifications.
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.
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
If two reports provide contradictory information:
Supervisor A:
Welding completed.
Supervisor B:
Welding 80% complete.
the system flags:
⚠️ Conflicting Progress Reports
Human verification required.
| 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.
The prototype dataset will consist of three primary components.
activity_id
wbs_level
activity_name
discipline
location
equipment
planned_start
planned_finish
dependencies
report_id
report_date
supervisor
discipline
raw_text
source
report_id
expected_activity_id
event_type
status
confidence_label
The system will be evaluated using labelled field-report/activity pairs.
- Top-1 Accuracy
- Top-3 Accuracy
- Precision
- Recall
- F1 Score
Evaluate extraction of:
- Activity
- Equipment
- Status
- Date
- Progress
- Location
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.
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.
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.
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
- PDF processing
- OCR
- Voice input
- Multilingual reports
- Advanced progress estimation
- Conflict detection
- Historical analytics
- Primavera/MS Project integration
- Enterprise PMIS integration
- Real-time notifications
- Project-level analytics
- Continuous terminology learning
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
"Which activities are behind schedule in Section B?"
The system could respond using the latest validated project data.
AI recommends; humans retain authority over important schedule changes.
Every AI-generated update should be traceable to its source report.
The AI cannot create arbitrary schedule IDs.
Only sufficiently reliable matches should be automatically processed.
Every mapping and modification should have:
Source
Timestamp
Activity ID
AI Decision
Confidence
Reviewer
Final Action
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
| Detail | Information |
|---|---|
| Problem Statement | SIH26122 |
| Theme | Smart Automation |
| Category | Software |
| Organization | Oil India Limited |
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.
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.
SIH26122 · Oil India Limited · Smart Automation