Data Analytics & Grounded AI Systems·Independent Analytics & AI Engineering Project

InsightDesk

A sales analytics dashboard with Python/Pandas calculations and an AI analyst grounded in computed summaries.

Built as a working showcase that can serve as a foundation for a customized business implementation.

Connect About This Project Source code

28 Tests

Automated Verification

28 passing pytest checks covering calculations, data hardening, session isolation, and AI grounding

$6.11M

Verified Revenue

$6,112,764.78 gross revenue across 1,144 unique orders in benchmark dataset

1,500 Rows

Benchmark Dataset

1,500 deterministic transaction records spanning 12 months (seed 42)

Zero Hallucination

Grounded AI Pipeline

Gemini receives pre-computed numerical summaries, never raw unverified CSV rows

Inspect the evidence

Review the implementation, data model and tests in the repository.

Source & README Commit history

MY CONTRIBUTION

What I built

Independent project

Deterministic sales aggregation engine computing revenue, orders, AOV, and discount rates
Monthly revenue trend analysis with month-over-month percentage growth calculation
Top product ranking with individual revenue contribution and unit volume shares
Regional sales breakdown across North America, EMEA, APAC, and LATAM
Automated rule-based business insight generator producing 3–5 key observations
Data quality and completeness profiler flagging missing values and inferred column types
Grounded natural language AI analyst powered by Gemini with verified context injection
Session isolation with bounded LRU cache preventing cross-dataset context contamination

01 / The Business Problem

Numerical Hallucination & Data Quality Failures in AI Dashboards

Prompting large language models with raw tabular records often leads to fabricated totals, faulty ranking logic, and credential leakage. Additionally, real-world CSV inputs contain formatting anomalies like currency symbols, commas, negative adjustments, and missing columns that break naive ingestion pipelines.

02 / The Solution

Two-Stage Architecture: 'Python Calculates; AI Explains'

InsightDesk enforces a strict two-stage pipeline. The Python/Pandas backend sanitizes inputs, preserves legitimate negative refund entries, computes exact metrics (revenue, AOV, monthly trends, regional share), and profiles schema completeness. The AI layer receives only structured analytical packets, and its answers undergo server-side validation against verified ground truth.

03 / PRODUCT EXPERIENCE

Core Capabilities & Workflow

MODULE 01

Deterministic Pandas Analytics Engine

Computes baseline executive KPIs (total revenue, unique order count, average order value, units sold, average discount), monthly revenue progression with month-over-month growth rates, top product rankings with revenue share, and regional volume distributions.

Practical Value:

Guarantees 100% mathematical consistency without relying on approximate probabilistic estimates.

Executive sales analytics dashboard displaying baseline KPIs, monthly revenue progression, and transaction telemetry

MODULE 02

Data Quality & Schema Profiler

Evaluates incoming datasets against a canonical sales schema, assessing completeness score, duplicate row detection, per-column null percentages, and data type inference. Sanitizes currency symbols, commas, and percentage representations while preserving negative refund values.

Practical Value:

Surfaces data hygiene issues transparently to the analyst prior to reporting.

Data quality profiler with healthy schema validation badge, completeness scoring, and column type breakdown

MODULE 03

Grounded Natural Language AI Analyst

Integrates Google Gemini to interpret analytical findings in natural language. Receives structured KPI contexts rather than raw rows, verifies cited metrics against backend ground truths, redacts API credentials on error, and gracefully informs when unconfigured.

Practical Value:

Provides executive-ready written explanations with zero risk of fabricated figures.

Grounded AI Analyst powered by Gemini answering natural language queries with verified numerical facts and follow-ups

MODULE 04

Bounded Session Isolation & In-Memory Cache

Generates unique session identifiers per loaded or uploaded dataset. Enforces an in-memory Least-Recently-Used (LRU) cache with a bounded capacity of 50 datasets to prevent memory exhaustion, returning HTTP 404 for unknown or expired identifiers.

Practical Value:

Prevents cross-dataset context leakage when querying multiple files sequentially.

Top products ranking and regional performance breakdown isolated by active dataset session context

04 / TECHNICAL ARCHITECTURE

Engineering Foundation & Architecture

Two-Stage Analytical Separation

Architected around the principle that numerical calculations belong in pure, vectorized Python/Pandas logic, while LLMs are strictly reserved for linguistic synthesis and contextual narrative.

Data Sanitization & Integrity Preservation

Implemented robust input sanitization stripping currency formatting and commas without silently clipping negative financial records such as returns and credit adjustments.

Server-Side Fact Verification Guardrails

Constructed verification logic comparing metrics cited in AI answers against computed values, rejecting ungrounded claims and protecting credentials across error boundaries.

Comprehensive Automated Test Verification

Engineered 28 pytest verification checks spanning mathematical ground truth, endpoint contracts, upload hardening, and dataset context isolation.

Technologies Used

FUNCTIONAL ROLE

Python 3.12

Core backend runtime for data manipulation and service endpoints

FastAPI & Pydantic

High-performance typed REST API and schema validation contracts

Pandas & NumPy

Deterministic vectorized financial aggregations and statistical profiling

React 19 & Vite

Reactive component interface with instant dashboard feedback

Recharts

Interactive visual analytics charts for monthly trends and product rankings

Google Gemini API

Grounded natural language insights with server-side fact verification

Pytest

Automated verification test suite across 28 functional test cases

05 / ENGINEERING DECISIONS

Technical decisions & trade-offs

DECISION 01

Vectorized Pandas over Raw LLM Arithmetic

Perform all math in Pandas before AI context compilation.

Why this approach

LLMs are probabilistic token predictors that cannot guarantee reproducible mathematical precision. Vectorized Pandas guarantees deterministic arithmetic.

DECISION 02

In-Memory LRU Cache over Persistent Database

Store active sessions in a bounded Python in-memory LRU cache with capacity 50.

Why this approach

Keeps the workstation lightweight and zero-infrastructure without running Docker, PostgreSQL, or Redis, while preventing unbounded memory growth.

06 / REFLECTIONS & LEARNINGS

What This Project Demonstrated

INSIGHT 01

Decoupling Computation from Narrative Generation

Designing InsightDesk proved that reliable AI applications require hard mathematical guardrails. By treating the language model as an explanatory interface rather than a calculation engine, data-driven systems remain both trustworthy and articulate.

INSPECT THE PROJECTSource, demo & implementation

Live Working Product

Deployed interactive web application on Vercel

Public Source Code

Read the implementation and browse its history.

Multi-Device Responsive

Inspect the layouts in the screenshots and demo.

Implementation details

Explore the architecture and decisions in this case study.

DATA ANALYTICS & GROUNDED AI