CA Hiring Automation Tool
Author : CA PRATEEK ATAL
Problem Statement
The hardest part of TDS compliance is not applying a known rate — it is identifying the cases that depend on the substance of a transaction rather than its surface. Most TDS tools available in the market work purely on the ledger classification already done by accountants; they inherit whatever section the bookkeeper assigned and never interrogate whether the underlying narration reveals a different substance. In practice, a meaningful share of vouchers are described only by a free-text narration, with no clean vendor name and a ledger head that may not reflect the true nature of the payment. Rule-based tools classify the well-structured entries correctly, but silently skip or mis-state these — understating TDS liability and creating downstream exposure to interest, disallowance under Section 40(a)(ia), and departmental notices.
Yet the obvious remedy — handing every classification to an AI — is unacceptable in a professional context, because a chartered accountant must be able to explain and defend every figure that reaches a return. The problem is therefore twofold: catch the TDS that rule-matching misses, and do so through a system where the AI advises while a deterministic, auditable engine computes — keeping the practitioner’s judgment, and accountability, central.
The Solution
The AI-Powered TDS Suite is a locally-run web application that pulls voucher data directly from Tally Prime and applies a two-stage classification pipeline. A deterministic rule engine first computes TDS for all clearly-structured entries. An AI advisory layer then resolves the cases the rules cannot — narration-only vendors and genuinely ambiguous expense sections — and surfaces its reasoning, confidence and the alternative section it considered, for the practitioner to review.
In the demonstration on a dummy dataset, the rule engine alone computes a baseline TDS figure; once the AI resolves the narration-only vendors that rule-matching could not read, the computed liability rises substantially — the difference representing TDS that a purely rule-based tool would have missed. Critically, the AI only suggests; the final TDS amount, thresholds and rates are always computed by the deterministic engine, so every number remains explainable and auditable.
Beyond classification, the Suite carries the work through to a return-ready output. It parses TDS challans and links them to the corresponding deductions, computes applicable interest and late fees, and exports the result in an Excel format that imports directly into the Winman TDS application — taking the practitioner from raw Tally data to filing-ready data within a single workflow.
Key Features
Stage 1 — Classification & TDS computation
Stage 2 — Reconciliation & return-ready output
Technology Used
| Layer | Technology |
| Frontend | Vanilla JavaScript (no framework), HTML5, CSS3 (CSS Grid) |
| Backend | Python with the Flask web framework; local REST API |
| Data source | Tally Prime via HTTP-XML gateway |
| AI / LLM | Google Gemini API— advisory layer only |
| Persistence | SQLite |
| Testing | pytest — 265 automated tests |
| Development | AI-assisted development using Claude Code (Anthropic) |
Relevance & Impact for the Profession
Every CA firm and finance team processing TDS faces the same blind spot: entries that the software cannot read are entries the firm cannot deduct. By combining deterministic computation with an AI advisory layer that explains itself, the tool directly addresses a recurring, high-cost compliance gap — while modelling a governance pattern (AI advises, the engine computes) that keeps the chartered accountant’s judgment, and professional accountability, firmly at the centre.