Transforming Construction Market Intelligence with AI-Driven Automation
Customer
Problem Statement
Business As-Is
The client has been delivering online market intelligence through multiple structured, layered queries. This process required significant user interaction, consuming both time and cost (in terms of storage and retrieval). Additionally, the system relied heavily on individual user memory and preferences to deliver insights.
Business To-Be
The client aimed to enhance its services by becoming a first mover in offering AI-enabled market intelligence. In today’s digital landscape, first-mover advantage is critical, it can often be a make-or-break factor for businesses.
Identified Business Use Cases
- Automated Data Presentation: Extract and present data from structured, semi-structured, scanned, and even handwritten business documents with minimal manual intervention.
- Natural Language Search: Enable non-digital-savvy users to perform context-aware searches in natural language and receive accurate results in near real-time.
- Document Summarization: Generate meaningful summaries of large tender and planning documents to help users make faster, informed decisions.
- Intelligent Communication: Quickly identify and trace business opportunities via email, enriched with relevant context to improve productivity and communication quality.
Solution
The client operated multiple database instances with several service layers built over decades using diverse technology stacks. Addressing this requirement demanded multiple tools, techniques, and efficient integration between components. Cost was another critical dimension, especially for “pay-as-you-use” services and cloud subscriptions.
Arrk recommended an iterative PoC strategy for rapid evaluation. During PoCs, we primarily leveraged NLP-based OpenAI models. Interestingly, different business outcomes required different tools. Ultimately, we selected OpenAI, Mistral, and models available on AWS Bedrock (including LLaMA and Claude). These foundation models were configured based on specific use cases.
The adopted tech stack included:
-
Python Flask API deployed on AWS ECS
-
pgvector as the vector database
-
Marimo for dashboarding and visualisation
Key challenges and how they were resolved:
-
Data Extraction from Scanned & Handwritten Forms: Initially, extracting data from scanned and handwritten application forms was a major challenge. We integrated MistralOCR after evaluating multiple methods, significantly improving text recognition accuracy.
-
Processing Large Document Volumes: We encountered gateway timeout issues during large-scale document processing. To resolve this, we implemented an asynchronous ingestion architecture using AWS Batch, enabling reliable and scalable processing.
As with any AI-based solution, accuracy and maturity depend on continuous training. We engaged in-house subject matter experts to extensively query and use the system, improving overall outcomes.
Outcomes
- Reduced manual operational effort by automating repetitive, time-consuming tasks.
- Significantly enhanced user experience and interaction quality.
- Accelerated deployment timelines, new AI use cases now go live within days instead of weeks.
- Improved consistency, accuracy, and reliability of responses through contextual data grounding and workflow orchestration.
- Established a foundational AI system that scales with expanding enterprise AI needs.
