TaxAssist — AI-Guided Income Tax Filing & Defective Return Fixer

A prototype targeting the Income Tax e-Filing Portal (incometax.gov.in).

Live demo: https://tax-assist-prototype.vercel.app/ Team: Nilesh Kumar Das, CSE, IIT Guwahati

---

1. The problem, in plain terms

Two real, evidence-backed pain points on India's Income Tax e-Filing Portal:

1. A citizen who receives a Section 139(9) "Defective Return" notice faces dense legal/technical jargon ("Schedule Part B-TTI", "e-Proceedings", "DIN") with zero plain-language explanation, and — if they agree there's a mistake — is forced to download a separate desktop "Offline Utility," manually rebuild their return, export a JSON file, and re-upload it. This is a process built for accountants, applied to ordinary citizens.
2. A first-time or simple filer faces the same underlying problem one step earlier: figuring out which ITR form applies, what each form requires, and how to compute what's owed — all before ever hitting a defect. Both problems currently push citizens toward paying a Chartered Accountant even for small, common corrections.

We picked this problem after an audit-first process, not a guess — see Section 3.

---

2. What we built

Two complete citizen journeys from one dashboard:

A. Fix a Defective Return (Section 139(9))

Notice → AI-translated plain-language summary (with the original legal text one tap away) → a short guided Q&A that replaces the offline-utility download entirely → simulated Aadhaar OTP verification → confirmation.

B. File a Fresh Return

A deterministic decision tree (plain-language questions, e.g. "Do you earn money mainly by offering a skill or service, or by running a business?") determines the correct ITR form (1/2/3/4) → a single reusable Q&A engine asks one question at a time for that form's required fields → a review screen → a simplified tax computation with an AI-generated plain-English breakdown → mock payment (only if tax is due) → OTP verification → a filed receipt with an acknowledgment number.

Every technical/legal term (80C, 80D, 44AD/44ADA/44AE, Chapter VI-A, TDS) is removed from the default question text and rephrased in plain English. A collapsed "Why we ask" toggle (reusing the same pattern as "View Official Notice Text") reveals the exact legal reference for anyone — a reviewing CA, a curious citizen — who wants it, without forcing jargon on everyone else.

---

3. How we got here — the full process

Phase 1 — Audit before building (Stage 1)

Before picking a specific flaw, we ran a structured audit of the real portal:

• Automated scans: Lighthouse and Pa11y across the home page, login page, and help/service pages (mobile, simulated slow 4G).
• Manual journey reconstruction: since we could not and should not log in with real credentials, we reconstructed the actual click-path for the 139(9) flow using public documentation, screenshots, and official user manuals — never touching a live government system with real data.
• Findings, evidence-backed, not assumed: a 38.7s Largest Contentful Paint on the login page under simulated slow 4G, 14 color-contrast failures, 59 duplicate ARIA IDs breaking screen readers, and — the flaw we chose to fix — a citizen journey that dead-ends at a mandatory desktop-software download for a correction that is often a one-line number.

Phase 2 — Problem selection, scored not guessed

We clustered every friction point found and scored each against what actually matters for this brief: universal pain, demo clarity, meaningful AI role, and mockability without real data. The jargon-translation-plus-offline-utility-replacement cluster scored highest and became our locked target — the "fix the defect" flow that shipped for Stage 1.

Phase 3 — Stage 2 expansion

After advancing past the Stage 1 shortlist, we identified a real gap in our own product thinking: fixing a defect alone still leaves the first-time filing process — the part that actually creates most CA dependency — untouched. If we only build the fixer, accountants simply use it to do their existing paid work faster; the citizen doesn't necessarily save money. We expanded to a second flow — fresh ITR filing — so the tool removes CA dependency across the journey, not just at the correction step.

---

4. Issues we hit, and exactly how we resolved them

We ran multiple independent audit passes (V1 through V5) against the live deployment throughout the build, specifically to catch problems before a judge would. Being upfront about what actually went wrong, and how, is itself part of the "Honesty" criterion this brief asks for.

This is the beginning of the README. Read the rest at the original. Read the full README ↗