How a leading Indian materials sciences company eliminated manual data compilation and replaced reactive reporting with a Conversational AI-powered financial intelligence solution, accelerating insight generation across its reporting workflows.
The finance team at a major Indian materials sciences company manages hundreds of SKUs and thousands of invoices. Leadership requires periodic, highly granular performance visibility—specifically deep-dives into Special Rate Contracts (SRC%) and Contribution Margins (CM).
Previously, the team built reports reactively. When leadership requested variance explanations, the team lacked the time to consolidate disconnected sales and cost data silos. They needed a system to replace manual spreadsheet management with integrated, on-demand insights.
Three problems driving cost and unpredictability:
Finance teams spent hours merging datasets, validating formulas, and reconciling reports instead of analyzing business performance.
Reporting relied on manually maintained Excel extracts, creating version-control issues, broken formulas, and dependency on individual analysts.
Reports were static summaries with limited drill-down capability, making variance analysis and KPI explanations slow and manual.
An intelligent financial layer bridging static reports and instant explainability.
A Conversational Analytics solution designed to ingest existing Excel files for SRC% and CM reports, transforming static data into an interactive intelligence layer without requiring a multi-year data warehouse migration.
Data Pipeline & Tech Stack
The frontend is built using React 18 (Vite) with a FastAPI (Python) backend. Structured data is managed in a PostgreSQL database via SQLAlchemy. The conversational agent framework utilizes LlamaIndex, powered by Claude Sonnet and Haiku via AWS Bedrock, with FAISS handling vector search.
The Intelligence Model
A semantic Knowledge Base allows the AI to understand specific KPI calculations, metrics, and business abbreviations. Users ask factual, natural-language questions, and the LLM queries the analytical data model to return precise tables. An “Expert Guidance” human-in-the-loop (HITL) mechanism continuously validates the AI’s logic to ensure absolute trust in the financial outputs.
A secure conversational portal deployed with role-based access for Admins and Business Users.
Admins upload Excel files and manage a central dictionary of KPI definitions, training examples, and business synonyms to govern the AI’s logic and maintain a single source of truth.
Users ask natural language questions with auto-suggested follow-ups to receive instant, fact-based tabular data. Every response includes a thumbs-up/down mechanism to capture direct user feedback.
Downvoted answers route directly to an Expert Review inbox. Domain experts correct the logic, continuously training the system and maintaining strict accuracy in the financial outputs.
Users can save generated charts into personalized dashboards or view standard uploaded reports using dynamic dimensional filters to track metrics over time.
Digitalized financial intelligence transformed reporting from a manual chore into an on-demand, conversational experience. By automating data traversal and calculation mapping, the finance team transitioned from data gatherers to strategic advisors.