Project: [Enterprise Agents] - [TaxMind]

[arorarajat9]

Track

Enterprise Agents (M365 Agents Toolkit)

Project Name

TaxMind

GitHub Username

@arorarajat9

Repository URL

https://github.com/arorarajat9/taxmind

Project Description

TaxMind helps Indian small businesses (kirana / retail stores) with GST filing — a task that normally costs days of manual Excel reconciliation and ₹2,000–5,000 per month in CA fees.

Upload a sales register, purchase register, and GSTR-2A and TaxMind:

1. Ingests messy Excel — misaligned columns, mixed date formats, Hindi or English headers — and maps it onto a GST business ontology.
2. Reconciles purchases against GSTR-2A (matched / mismatched / missing) and computes ITC at risk.
3. Flags blocked ITC under Section 17(5) — and every flag carries the exact cited GST Act section + snippet retrieved from a knowledge base (e.g. catering → 17(5)(b)(i), construction → 17(5)(d)). No citation, no assertion; low confidence → "consult a CA".
4. Detects anomalies — duplicates, missing/invalid GSTINs, future-dated invoices.
5. Generates GSTR-1 & GSTR-3B summaries plus a filing-ready Excel workbook.
6. Keeps a human in the loop — it prepares filings, never auto-submits to GSTN. Every decision lands in a cited audit trail.

How the Microsoft IQ stack is the hero:

• Foundry IQ (Azure AI Search agentic retrieval) grounds every compliance flag in the cited GST Act text.
• A Fabric IQ–style typed ontology lets the agent reason in GST language (ITCClaim, TaxLiability, Invoice, Supplier) instead of raw columns.
• The Microsoft Agent Framework orchestrates four specialist agents: reconcile → ITC eligibility → anomaly → return generation.

Built with a provider abstraction: runs fully offline in local mode (bundled public GST Act text) for reproducible evaluation, and upgrades to live Azure / Foundry IQ via a single .env with zero code changes.

All demo data is synthetic; only public GST Act / CBIC material is used.

Demo Video or Screenshots

Dashboard screenshot: https://github.com/arorarajat9/taxmind/blob/main/demo/assets/dashboard.png
Walkthrough: https://github.com/arorarajat9/taxmind/blob/main/demo/walkthrough.md
Demo video script: https://github.com/arorarajat9/taxmind/blob/main/demo/video-script.md

Primary Programming Language

Python

Key Technologies Used

• Microsoft Agent Framework (multi-agent orchestration: reconcile → ITC eligibility → anomaly → return generation)
• Foundry IQ / Azure AI Search (agentic retrieval — grounds every flag in cited GST Act text)
• Fabric IQ–style typed ontology (Invoice, Supplier, PurchaseEntry, ITCClaim, TaxLiability via pydantic)
• Azure OpenAI (GPT-4o) for LLM-assisted column mapping (with deterministic rule-based fallback)
• FastAPI backend, Next.js 14 + Tailwind + Recharts frontend
• pandas / openpyxl for messy-Excel ingestion; Azure Blob Storage for knowledge sources
• GitHub Copilot in VS Code

Submission Type

Individual

Team Members

No response

Submission Requirements

• My project meets the track-specific challenge requirements
• My repository includes a comprehensive README.md with setup instructions
• My code does not contain hardcoded API keys or secrets
• I have included demo materials (video or screenshots)
• My project is my own work with proper attribution for any third-party code
• I agree to the Code of Conduct
• I have read and agree to the Disclaimer
• My submission does NOT contain any confidential, proprietary, or sensitive information
• I confirm I have the rights to submit this content and grant the necessary licenses

Quick Setup Summary

TaxMind runs fully on localhost — no Azure account, no API keys required.

Prerequisites: Python 3.11+ and Node.js 18+

make setup # one-time: Python venv + deps, frontend deps, and synthetic kirana demo data
make demo # runs the full pipeline in the terminal — reconciliation, cited
# Section 17(5) blocked-ITC flags, GSTR-1/3B summaries, and a
# filing-ready Excel workbook written to ./output/
make dev # starts FastAPI backend + Next.js dashboard together → http://localhost:3000
make test # acceptance test suite

No make? See the manual steps in README.md (pip install -r requirements.txt → generate data → run pipeline).

Optional Azure / Foundry IQ upgrade: copy .env.example to .env, fill in your Azure
OpenAI + AI Search + Foundry IQ values, and the app auto-detects and switches to
live Foundry IQ agentic retrieval — zero code changes. Full instructions in README.md
and docs/azure-setup.md.

Technical Highlights

• Grounded compliance, not guesses: every blocked-ITC flag carries the exact cited GST Act section + snippet retrieved from Foundry IQ (e.g. catering → 17(5)(b)(i), construction → 17(5)(d)). When retrieval confidence is low, the agent says "not certain — consult a CA" instead of hallucinating a rule.
• Provider abstraction: one interface, two backends auto-selected from .env. Local mode (bundled public GST Act + deterministic reasoning) makes the demo reproducible with zero keys; Azure mode swaps in real Foundry IQ + Azure OpenAI with no code changes. The same cited {section, snippet, source} shape flows through either path.
• Reasons in GST language via a Fabric IQ–style typed ontology (ITCClaim, TaxLiability) rather than raw spreadsheet columns.
• Robust messy-Excel ingestion: header-row detection, mixed date formats, Hindi/English headers, NaN-safe GSTIN handling so blank GSTINs don't misclassify B2C sales as B2B.
• Human-in-the-loop by design: prepares filings + a cited audit trail, never auto-submits to GSTN.

Challenges & Learnings

• Grounding over fluency: the hard part wasn't generating answers but refusing to assert anything without a citation, and degrading gracefully (fall back to "consult a CA") when retrieval confidence is low.
• Building for two runtimes: designing a single interface that runs identically offline (for reproducible evaluation) and on live Azure / Foundry IQ taught me a lot about clean provider abstraction.
• GST domain nuance: e.g. food/catering is blocked under 17(5) but resale food inventory keeps its ITC — encoding that distinction correctly mattered for real-world accuracy.
• Real Excel is hostile: misaligned columns, mixed locales, and NaN values required defensive parsing to avoid silent misclassification.

Contact Information

rajatarora9645@gmail.com

Country/Region

India