TaxPilot

Transparent UK tax advisor with evidence-backed calculations

TaxPilot's goal is to turn UK tax advice from a black-box chat answer into a traceable decision system. It starts by identifying the tax year, residency, domicile, income, gains, reliefs, and missing facts, then routes the question through official sources, structured knowledge, and deterministic calculators before an answer is shown.

The larger ambition is advisory, not just form filing: a user should be able to ask about income, capital gains, residency, remittance, reliefs, or cross-tax planning and see which rules were used, which numbers came from calculators, where the authority sits, and which parts remain uncertain.

Trust model

ClassifyDecompose the question into tax domains, topics, and per-fact issues.
ResearchParallel agents research each topic against legislation and HMRC manuals via hybrid retrieval.
Review & decideA continuous agent reads findings, calls more tools until satisfied, and writes structured decisions.
Synth + sandboxDecisions become JavaScript executed in a vm2 sandbox — numbers come from code, not model arithmetic.
AdviseAn advisor writes the final note: main scenario plus alternatives, with uncertainty explicitly flagged.

The trust model splits responsibility deliberately. The language model handles intake, explanation, and orchestration; the knowledge store owns source-grounded rules; deterministic code owns numbers; and the uncertainty layer marks gaps, stale law, or model disagreement instead of turning them into confident prose.

Knowledge store

The knowledge layer treats tax law as structured, dated knowledge rather than loose text. A knowledge unit can carry statutory references, HMRC manual links, rates placeholders, effective dates, exceptions, examples, and cross-tax relationships, so retrieval can pull back both the relevant rule and the authority behind it.

Part Current fact Trust role
Structured knowledge units Stored in Postgres. Keeps conclusions tied to explicit knowledge units.
pgvector Supports hybrid retrieval. Connects questions to relevant sources.
Regulations / HMRC sources Used as authoritative trace-back material. Prevents model-only conclusions.

Engineering progress

The current build has a deterministic calculation skeleton with 85 tests and an agentic advisor that can already produce officially correct tax figures end-to-end through its tool loop. The next engineering frontier is widening coverage while keeping every number, citation, and missing-calculator gap visible.

Area Status
Deterministic calculation Self-built rules-engine skeleton with 85 tests.
Agentic advisor Vercel AI SDK multi-model tool loop; already able to produce officially correct tax figures end-to-end.
Tech stack Next.js, Vercel AI SDK, Postgres, Prisma.

Development approach

TaxPilot is being built with an OpenCode multi-agent team: manager for coordination, distiller for knowledge distillation, builder for implementation, and verifier for verification.

Current status

Status: in progress. TaxPilot is not a finished product, not live as a product, and not production-ready.

What exists today is the trust spine: deterministic calculation, source traceability, explicit cannot-answer flags, the Postgres + pgvector knowledge store, and the OpenCode multi-agent development approach.