Agentic Indian income-tax advisor

Tax answers you can actually trust — computed, cited, and refused when out of scope.

Most “AI tax” tools let a language model do the arithmetic and hope. TaxBrainAI hands every number to a deterministic engine, grounds every claim in the statute, and runs a fail-closed verifier that would rather refuse than guess.

9-node agent graph + HITL8-stage hybrid + graph RAGDeterministic compute engine139 engine tests + full backend suite

Live demo · try it

See it reason — then run it yourself.

First a guided walkthrough of the agent — routing, retrieval, a human-approval gate, a cited answer. Then the deterministic tools below, running live in your browser: no login, no server call, matched to the engine to the rupee.

TaxBrainAI agentguided demo

Pick a question below to watch the full pipeline light up stage by stage — routing, retrieval, the deterministic engine, the human-approval gate, then a cited answer. Or type your own at the bottom — it runs against the live backend.

Checking live-agent access…

See the live agent

Request access

The full agent — live retrieval, the model, and document tools — runs behind auth. Email to request access and I’ll set you up with a walkthrough, usually within 24 hours.

Request access by email

Replies usually within 24 hours.

What you’re watching

  • Routed & retrieved. The agent classifies intent, then runs the 8-stage statute retrieval.
  • Human-approved. Anything it would act on pauses for an explicit yes/no.
  • Cited & computed. Figures come from the deterministic engine; every claim carries its section.
  • English or हिंदी. Ask in either — it answers in the question’s language, citing the same sections.
  • Or refused. Out-of-scope questions are declined, not guessed.

Scripted sessions that mirror the live agent (which runs against the FastAPI backend, behind auth).

live now

Run the tools yourself — no login

These are the exact deterministic rules the agent calls, ported to run in your browser and checked against the engine’s test suite — so they cost nothing and work even if the server is down.

Income tax · AY 2026-27 · new vs old

Income

Capital gains & VDA

Deductions · old regime

You

AY 2026-27. Deductions and HRA apply under the old regime only (80CCD(2) also under the new). Everything else the engine takes is here; three specialist paths — the Section 112 land/building indexed lower-of, presumptive Sections 44AD/44ADA, and AMT (Section 115JC) — run in the full engine, open-sourced as taxbrainai-compute.

New regime

recommended
₹97,500
total tax payable
Taxable income (normal)₹14,25,000
Tax on slabs₹93,750
Health & education cess (4%)₹3,750
Total tax₹97,500

Saves ₹1,59,900 versus the old regime.

Old regime

₹2,57,400
total tax payable
Taxable income (normal)₹14,50,000
Tax on slabs₹2,47,500
Health & education cess (4%)₹9,900
Total tax₹2,57,400

Computed in your browser — the same deterministic logic the agent calls, differential-tested against the Python engine to the rupee.

Statute lookup & 1961 → 2025 mapping

Search any section by number or by what it does — the same resolver the agent uses to turn a reference into a plain-English citation. Where the Income-tax Act, 2025 renumbers a section, the row also shows its new number, straight from the statute graph’s 1961↔2025 equivalence edges.

55 of 55 sections2025 Act →rows show the new Act’s number where the statute graph has a 1961↔2025 equivalence edge
  • 115BACNew tax regime (default)

    The default regime from AY 2024-25. Lower slab rates, but most deductions and exemptions no longer apply.

  • 87ARebate for small incomes

    New regime: up to ₹60,000 when total income is ≤ ₹12,00,000. Old regime: up to ₹12,500 when ≤ ₹5,00,000. Residents only.

    2025 Act → Section 156 — Rebate of income-tax

  • 16(ia)Standard deduction (salary)

    A flat deduction from salary — ₹75,000 under the new regime, ₹50,000 under the old.

    2025 Act → Section 19 — Deductions from salaries

  • 10(13A)House Rent Allowance (HRA)

    Exempts the least of: actual HRA, rent paid minus 10% of salary, or 50% of salary (metro) / 40% (non-metro). Old regime only.

    2025 Act → Section 11 — Incomes not included in total income

  • 10(10D)Life-insurance proceeds

    Maturity or death proceeds are exempt, subject to premium-to-cover limits — high-premium ULIPs (premium above ₹2,50,000) and other policies (above ₹5,00,000) are taxable.

    2025 Act → Section 11 — Incomes not included in total income

  • 24(b)Home-loan interest

    Interest on a housing loan. Self-occupied: up to ₹2,00,000 a year (old regime). New regime allows no interest on a self-occupied home; on a let-out home it is allowed but the loss can't be set off against other income.

    2025 Act → Section 22 — Deductions from income from house property

  • 80CInvestment deductions

    Up to ₹1,50,000 for PPF, ELSS, life insurance, principal repayment, and more. Old regime only.

  • 80CCD(1B)Extra NPS deduction

    An additional ₹50,000 for NPS contributions, over and above the ₹1,50,000 Section 80C ceiling. Old regime only.

  • 80CCD(2)Employer NPS contribution

    The employer's NPS contribution — the one Chapter VI-A deduction still allowed under the new regime.

  • 80DHealth insurance premium

    Premium paid for self, family, and parents. Old regime only.

    2025 Act → Section 126 — Health insurance premia

  • 80DDDisabled-dependant care

    Maintenance and medical care of a disabled dependant — ₹75,000, or ₹1,25,000 for severe disability. A flat amount, regardless of spend. Old regime only.

    2025 Act → Section 127 — Maintenance of a disabled dependant

  • 80DDBTreatment of specified diseases

    Up to ₹40,000 for prescribed serious illnesses, or ₹1,00,000 for a senior citizen. Old regime only.

    2025 Act → Section 128 — Medical treatment

  • 80UDisability (self)

    For a resident with a certified disability — ₹75,000, or ₹1,25,000 for severe disability. Old regime only.

    2025 Act → Section 154 — Person with disability

  • 80EEducation-loan interest

    The full interest on a higher-education loan, for up to 8 years. No upper limit; old regime only.

  • 80GDonations

    Donations to approved funds and charities — 50% or 100%, some capped at 10% of income. Old regime only.

    2025 Act → Section 133 — Donations to certain funds

  • 80TTASavings-account interest

    Interest on savings deposits, up to ₹10,000. Old regime; not for senior citizens, who use 80TTB.

  • 80TTBSenior-citizen deposit interest

    For senior citizens — interest on savings and fixed deposits, up to ₹50,000. Old regime only.

  • 111AShort-term capital gains · listed equity

    20% on STCG from listed shares and equity funds, for transfers on or after 23 July 2024.

    2025 Act → Section 196 — Tax on short-term capital gains

  • 112ALong-term capital gains · listed equity

    12.5% on LTCG above ₹1,25,000 a year from listed shares and equity funds.

    2025 Act → Section 198 — Tax on long-term capital gains, certain cases

  • 112Long-term capital gains · other assets

    12.5%. Land or building bought before 23 July 2024 may take the lower of 12.5% unindexed or 20% indexed.

    2025 Act → Section 197 — Tax on long-term capital gains

  • 54Rollover · sale of a house

    Long-term gain on a residential house is exempt to the extent you buy or build another residential house within the prescribed windows.

    2025 Act → Section 82 — Profit on sale of property used for residence

  • 54FRollover · other asset into a house

    Long-term gain on any asset other than a house is exempt if the net sale proceeds go into one residential house — proportionate if only part is invested.

  • 54ECRollover · capital-gains bonds

    Long-term gain on land or a building is exempt if invested in NHAI / REC-type bonds within 6 months — capped at ₹50,00,000 a year.

  • 115BBEUnexplained income

    Unexplained cash credits, investments or money (Sections 68–69D) are taxed at a flat 60%, plus surcharge and cess — no deduction, allowance or loss set-off allowed.

  • 115BBHVirtual digital assets

    A flat 30% on crypto and other VDA income. Only cost of acquisition is deductible; no loss set-off and no 87A rebate.

  • 194STDS on VDA transfer

    1% tax deducted at source on the transfer of a virtual digital asset.

  • 44ADPresumptive · business

    Income deemed at 6% of digital turnover and 8% of cash turnover. Turnover cap ₹2 crore, or ₹3 crore when cash receipts are ≤ 5%.

  • 44ADAPresumptive · profession

    Income deemed at 50% of gross receipts. Receipts cap ₹50 lakh, or ₹75 lakh when cash receipts are ≤ 5%.

  • 44ABTax audit

    Accounts must be audited above ₹1 crore turnover (₹10 crore if cash is ≤ 5%) for business, or ₹50 lakh receipts for a profession.

  • 194ATDS · interest

    Tax deducted at source on interest other than on securities — e.g. bank or company deposit interest.

  • 194CTDS · contractors

    Tax deducted at source on payments to contractors and sub-contractors.

  • 194HTDS · commission / brokerage

    Tax deducted at source on commission or brokerage.

  • 194JTDS · professional fees

    Tax deducted at source on fees for professional or technical services, and on royalty.

  • 194TTDS · payments to partners

    Tax deducted at source on a firm's salary, remuneration, bonus, commission or interest paid to its partners. New from 1 April 2025.

  • 139(1)Filing the return

    The core filing obligation and its due dates — 31 July for most individuals, 31 October where an audit applies.

    2025 Act → Section 263 — Return of income

  • 139(4)Belated return

    A return filed after the due date, up to 31 December of the assessment year.

    2025 Act → Section 263 — Return of income

  • 139(5)Revised return

    Correcting an omission or error in a filed return, up to 31 December of the assessment year.

    2025 Act → Section 263 — Return of income

  • 139(8A)Updated return (ITR-U)

    File with additional tax up to 48 months after the end of the assessment year.

    2025 Act → Section 263 — Return of income

  • 139(9)Defective return

    Your return is flagged as defective, with time given to correct it.

    2025 Act → Section 263 — Return of income

  • 234AInterest · late filing

    1% a month on unpaid tax, from the due date to the date of filing.

    2025 Act → Section 423 — Interest, default in furnishing return

  • 234BInterest · advance-tax shortfall

    1% a month on the shortfall when advance tax paid is too little.

    2025 Act → Section 424 — Interest, default in advance tax

  • 234CInterest · deferred advance tax

    1% a month for missing the quarterly advance-tax installment deadlines.

    2025 Act → Section 425 — Interest, deferment of advance tax

  • 234FLate-filing fee

    A flat ₹5,000, reduced to ₹1,000 when total income is ≤ ₹5,00,000.

    2025 Act → Section 428 — Fee, default in furnishing return

  • 89Relief on arrears

    Relief when salary arrears or advance push you into a higher bracket — recomputed by spreading the income across the relevant years (Form 10E).

    2025 Act → Section 157 — Relief for salary arrears

  • 90DTAA relief

    Relief from double taxation under a tax treaty with the other country — by exemption or foreign-tax credit.

  • 91Foreign-tax relief (no treaty)

    Unilateral credit for foreign tax paid where India has no treaty with that country.

    2025 Act → Section 160 — Countries with no agreement

  • 143(1)Intimation after processing

    An automated summary after your return is processed — a refund, a demand, or no change.

    2025 Act → Section 270 — Assessment

  • 142(1)Inquiry before assessment

    A notice asking for information, or to file a return, before assessment.

    2025 Act → Section 268 — Inquiry before assessment

  • 148Income escaping assessment

    A reassessment notice when the department believes income was not fully disclosed.

    2025 Act → Section 280 — Income escaping assessment

  • 148AProcedure before a 148 notice

    The mandatory inquiry and opportunity to explain before a Section 148 reassessment notice can be issued.

    2025 Act → Section 281 — Procedure before a Section 280 notice

  • 144BFaceless assessment

    The faceless, jurisdiction-free assessment procedure.

    2025 Act → Section 273 — Faceless assessment

  • 156Notice of demand

    A formal demand for tax, interest, or penalty payable.

    2025 Act → Section 289 — Notice of demand

  • 245Adjustment of refund

    Notice that a refund is being set off against an outstanding demand.

    2025 Act → Section 438 — Set off and withholding of refunds

  • 133(6)Call for information

    The department may require information relevant to an inquiry or proceeding.

    2025 Act → Section 252 — Power to call for information

  • 288A / 288BStatutory rounding

    Total income (288A) and tax payable (288B) are each rounded to the nearest ₹10.

What it does

Not a chatbot with a tax prompt.

A compute engine, a knowledge graph, and guardrails it can’t talk its way around.

Deterministic tax math

AY 2026-27 new-regime marginal relief, the Section 87A rebate, the July 2024 capital-gains cutover, indexed long-term gains, surcharge caps, and the 30% rate on virtual digital assets — computed exactly, not estimated.

Statute-grounded answers

Every figure and rule cites its section. The answer carries its receipts, so you — or your CA — can check it line by line.

Graph-RAG over cross-references

Indian tax law is a web of cross-references. A graph stage follows the section-to-section links that flat vector search alone would miss.

Refuses out of scope

Ask it something it can’t ground and it declines instead of hallucinating. “Never fabricate” is wired into the verifier, not left to the prompt.

Human-in-the-loop on notices

Notice handling pauses at a LangGraph interrupt for human review, then resumes idempotently from its checkpoint.

PII-safe by construction

PAN, Aadhaar, names, and identifiers are graded and redacted before they reach the model or the logs, with an audit trail on every request.

How it works · architecture

Two coupled systems — the runtime graph and the build lifecycle.

Two systems, one hand-off: the nine-node request graph — where the engine numbers and the statute citations converge at the Synthesizer — and the separate, offline build lifecycle that produces the fine-tuned model driving it. Click any node to expand its detail.

live / verified explored · measured, SFT shipped model-driven
runtime ·  01 / 05  ·  Router
Runtime — one request flows through01 → 05 · Doc (when documents attached) → Retrieve · Compute → Planner / Notice → HITL → Synthesizer
01Routerrules gate02RetrieveBM25 · BGE-M303ComputeDecimal04Synthesizercite · transcribe05Verifyallow-list · hash-chainDoc26AS ↔ AISPlannerold vs newNotice143(1)·148HITLinterrupt()
✦ fine-tuned brain · Gemma-4-E4B
produced by the build lifecycle ↓ · in tool-serving mode drives ✦ Router (turn-1 action) + ✦ Synthesizer (routine compose) ↑
Build lifecycle — offline, produces the brainB1 → B4 · SFT ships · DPO / GRPO explored off SFT, head-to-head
B1datacorpusB2SFTLoRAB3eval366-rowB4servevLLM · bf16DPOpreferenceGRPOreward
Runtime — one request flows down
Conditional branches · human-gated → Synthesizer
✦ fine-tuned brain · Gemma-4-E4B
produced by the build lifecycle ↓ · in tool-serving mode drives ✦ Router (turn-1 action) + ✦ Synthesizer (routine compose) ↑
Build lifecycle — produces the brain
Explored off SFT — head-to-head, not sequential
…then the ship path continues
runtime · 01live

Router — deterministic scope + routing

Classifies intent (qa · compute · planning · notice · reconcile · filing) and routes to tools. Scope safety runs first and is never delegated to the answering model: out-of-scope is refused by deterministic rules; jailbreak/unsafe is screened by a dedicated safety classifier (keyword guard offline · Llama Guard 4 live) — before anything runs.

Intent classified, scope checked by rules — not the model.

The stack behind it

Ports & adapters everywhere: the core runs and is fully tested on deterministic mocks with zero API spend, then swaps to live providers behind one flag — numpy vectors → pgvector, a networkx graph → Neo4j, an in-memory checkpointer → a durable store, each behind the same port.

Frontend

  • Next.js App Router / RSC
  • Cloudflare Workers (OpenNext)
  • Streaming over SSE
  • Tailwind v4

Backend

  • FastAPI · REST + SSE
  • LangGraph · 9-node + HITL
  • FastMCP · 8 tools
  • Checkpointed HITL

Retrieval

  • BM25 lexical
  • Dense vectors · numpy → pgvector
  • Graph · networkx → Neo4j
  • RRF + cross-encoder rerank

Reliability

  • Deterministic Decimal engine
  • Gated eval harness
  • PII + audit guards
  • Ports & adapters

Measured & gated

The held-out numbers, and the gates around them.

A 200-question golden set drives development; the measured figures below are the run on a 366-row decontaminated held-out set (paired-bootstrap CI, Holm-corrected). The gate figures are hard pass/fail thresholds the run must clear — they are not accuracy scores. Graded axes are scored by the real-tier LLM-judge ensemble (two graders + an auditor).

0.90
overallmeasured
0.87
groundingmeasured
0.76
citation F1measured
0.88
hard-case citationmeasured
0.99
faithfulnessmeasured
0.99
format validitymeasured
0.87
Hindimeasured
0
PII leakagemeasured
100%
compute exactnessgate
100%
notice classificationgate
100%
jailbreak resistancegate
cites Section 115BBH → resolved · maps across the 1961 and 2025 Acts

Engineering

Decisions, and what they cost.

Every choice here buys something and pays for something. These are the ones I made on purpose.

Deterministic compute, not LLM arithmetic

Why
Tax math must be exact and auditable. A model that’s right 99% of the time is wrong on someone’s return.
Trade-off
We maintain a real rules engine with its own property tests instead of leaning on the model.

Fail-closed verifier

Why
An uncited number is treated as a bug — the verifier blocks any claim it can’t trace to a retrieved source.
Trade-off
It will occasionally refuse rather than answer confidently-but-wrong.

Hybrid + graph retrieval

Why
Statutes cross-reference constantly; lexical, dense, and graph retrieval each catch what the others drop.
Trade-off
Eight stages cost latency, so the pipeline runs against a strict time budget.

Ports & adapters (mock ↔ real)

Why
The whole system runs and is fully tested with zero API spend, then swaps to live providers behind one flag.
Trade-off
Every dependency needs a clean interface and a mock — more boilerplate, far more testability.

Cost & performance

Free where it can be, metered where it must be.

No invoice to quote — the interesting part is how it's engineered to stay cheap. Six decisions do the work.

Deterministic engine, not a metered model

live

Every tax figure is computed in code, never billed per token by an LLM. The arithmetic carries no per-query cost — and it's exact.

The tools run in your browser

live

The calculator and statute lookup are the engine ported to client-side code. They execute on your device, so they add no server cost at all.

Scale-to-zero model serving

live

The model tier serves on demand with min-containers at zero: when no one is asking, nothing is running, so idle time bills nothing.

Auth-gated GPU

live

The model endpoint sits behind proxy-auth, so unauthenticated traffic can't spin up a GPU. Abuse can't quietly run up the bill.

A fine-tuned small model

live

A fine-tuned Gemma-4-E4B, served on demand, stands in for a frontier API on every call — capability where it's needed, without the per-token frontier price.

Hard ceilings, cheap cold starts

live

A max-containers cap bounds the worst case, and enforce-eager serving trims cold-start cost — so spend can't run away.

The deterministic tools are live and free on this page; the model tier serves behind auth with scale-to-zero ceilings — engineered to keep spend near-zero, not a running invoice.

running

Free-live · no login

  • Deterministic tax computation (slabs, rebate, surcharge, cess)
  • Old-vs-new regime comparison
  • Section lookup and citation
  • 1961-to-2025 section mapping
running

Request access · auth-gated

  • The full agentic chat (graph + model)
  • Notice draft-assist
  • Document reconciliation
  • Vision extraction from uploads

Scorecard

What’s built, in eight numbers.

1,488
statute sections
4,707
graph edges
9
agent nodes
8
retrieval stages
8
MCP tools
200
golden questions
139
engine tests · 100% cov
92%
backend coverage · 353 tests

Graph and test counts are read straight from the repository. The agent graph runs end to end on deterministic adapters today; the live model tier swaps in behind one flag.

Statute graph

built

1,488 sections and 4,707 edges, built from the Income-tax Acts of 1961 and 2025.

Statute & rules corpus

built

The Acts, the Income-tax Rules, Finance Acts, and CBDT circulars — chunked and indexed for retrieval.

1961-to-2025 mapping

built

Equivalence edges linking each old section to its counterpart in the new Act.

Case law

built

309 full-text judgments fetched from Indian Kanoon — court, date, and citation captured for each — forming the case-law corpus beside the statute index.

Tested, cited, and verifiable end to end.

Property-tested, statute-grounded, and accurate to the rupee — and the load-bearing pieces are published, so you can verify them yourself: install the engine, plug in the MCP tools, or inspect the model card.

taxbrainai-compute

PyPI · MIT

The deterministic engine behind this page, published as a standalone library — pure-Decimal, zero dependencies, 139 tests at 100% coverage (20 Hypothesis property invariants), cross-verified against all four Income-Tax Department e-filing utilities.

taxbrainai-mcp

Official MCP Registry · MIT

The engine exposed as 8 tools over the Model Context Protocol — deterministic tax computation and statute lookup in any MCP client. Published to PyPI and listed on the official MCP Registry as io.github.harshil-projects/taxbrainai-mcp.

taxbrainai-gemma-4-e4b

Hugging Face · Apache-2.0

The fine-tuned advisor-synthesis model (Gemma 4 E4B) that drives the Router and Synthesizer — released with the merged bf16 ship weights, all three LoRA adapters (SFT · DPO · GRPO), and a full training-and-eval model card.

Try the live demoCalculator · liveStatute lookup · liveDeterministic & cited200-question eval set

The builder

Want this kind of engineering on your team?

TaxBrainAI is my work end to end — the agent graph, the deterministic engine, the fine-tune and its eval harness, and this page. I’m open to agentic-AI and backend roles.

TaxBrainAI — by Harshil

Agentic Indian income-tax advisor — deterministic compute, statute-grounded answers, and a fail-closed verifier.

Educational demo — general guidance, not personal tax, legal, or financial advice. Verify with a qualified professional before filing.

© Harshil · Built with Next.js, FastAPI & LangGraph.harshilprojects.com@harshil-projectsharshilprojects3@gmail.comtaxbrainai-compute · PyPItaxbrainai-mcp · MCP Registry