Indian tax returns from the terminal, because LLMs can't do math
To split tax filing the right way: LLMs read the documents, deterministic Python does the math, and you alone click Submit.

What it does
itr-wala prepares Indian income-tax returns for AY 2026-27 entirely on your machine. You hand your Form 16, AIS JSON, and broker P&L to an AI agent; the agent transcribes figures into a strict schema, and a deterministic Python engine computes slabs, rebates, surcharge, cess, and interest under both regimes. The result is a field-by-field filing pack and a regime comparison, while you retain sole control of Pay, Submit, and e-Verify on the government portal.
The interesting bit
The author treats the Income Tax Act as a specification suite: 47 “golden tests” are hand-derived from statute to pin every edge case—87A marginal relief, capital-gains exemption ordering, surcharge tiers, belated-return regime lockouts—while a property-based fuzzer sweeps randomized returns to enforce legal invariants like “more income never means less tax.” It is essentially a formal-methods attitude applied to personal finance, with the LLM relegated to the only job it is trusted with: reading PDFs and talking to you.
Key highlights
- Deterministic tax engine with 47 hand-derived golden tests and a 104-test input validator; CI runs 3,000 fuzz cases on every commit.
- Covers both regimes, salary, house property, capital gains (111A/112A/112), crypto/VDA, lottery, dividends, family pension, Chapter VI-A deductions, and advance-tax interest to the challan date.
- Zero network calls during install and zero pip dependencies; runs offline on Python 3.9+ stdlib only.
- Validator aggressively rejects malformed inputs and PAN/Aadhaar-shaped strings before they reach the engine.
- Honest scope boundaries: refuses non-residents, F&O, audit cases, foreign tax credit, and ESOP deferral rather than silently approximating.
Caveats
- Document transcription is still handled by the LLM, so a misread digit is the residual risk; the tool mitigates this by cross-checking Form 16 against 26AS/AIS totals, but the README is explicit that this is the weakest link.
- Native Windows support is on the roadmap; WSL works today.
- Rates are pinned to AY 2026-27; the engine refuses to compute other years rather than silently apply stale slabs.
Verdict
Use it if you file Indian taxes yourself, distrust black-box software, or want a reproducible computation that treats the statute as a spec. Skip it if you are a non-resident, trade F&O, or want someone else to click Submit for you.
Frequently asked
- What is karanb192/itr-wala?
- To split tax filing the right way: LLMs read the documents, deterministic Python does the math, and you alone click Submit.
- Is itr-wala open source?
- Yes — karanb192/itr-wala is an open-source project tracked on heatdrop.
- What language is itr-wala written in?
- karanb192/itr-wala is primarily written in Python.
- How popular is itr-wala?
- karanb192/itr-wala has 650 stars on GitHub.
- Where can I find itr-wala?
- karanb192/itr-wala is on GitHub at https://github.com/karanb192/itr-wala.