The incumbent sells advisors a tool that produces a static PDF. Lumio centers the client — handing them an interactive Planning Playground where they drag a Roth-conversion dial and watch brackets, cliffs, and IRMAA tiers move in real time — powered by an engine that doesn't just recompute scenarios but solves them: "the highest conversion that stays under your next IRMAA tier is $47,300."
Everything is a wrapper around one deterministic function —
compute_tax(facts, year, jurisdiction, assumptions). The differentiated surface is the
interactive client explorer sitting on top of an automated optimization layer.
Deterministic, auditable, versioned by tax year. Pure function, exact Decimal math,
a line-level trace that explains every number. Everything depends on it; it depends on nothing.
Not "recompute a scenario" but find the optimal input under a constraint — plus the response-surface sweep and breakpoint enumeration that make the client's sliders smart instead of dumb.
◆ DifferentiatorAn automated report plus a gamified sandbox the client explores. Live knobs → live surface; breakpoints light up; honest captions ("$12,000 of room left before the 24% bracket").
◆ DifferentiatorOCR a finished 1040 or IRS transcript into the baseline in seconds, with confidence scoring and human-in-the-loop review. Advisors will not hand-type returns at scale.
Roll forward N years for lifetime tax / after-tax-wealth comparison (Roth ladders) — but with the discount, growth, and inflation assumptions made explicit and advisor-tunable.
Roster, documents, comms, engagement letters, §7216 consent, IRS-notice tracking — so the prep engine (ProConnect / Drake) becomes a swap-out component, not a foundational dependency.
Scenario modeling and lifetime projection are not differentiators; Holistiplan already markets both. The honest wedge is the four things it does not do — and a client surface it has never had.
The broad market — active CPAs, EAs, PTIN-only preparers, and inactive-CPA practitioners. The advisor is the operator and the buyer.
The end consumer receives the report and the live Playground. Saved scenarios route back to the advisor — turning an annual chore into an ongoing conversation.
The incumbent just raised prices. A challenger enters on price, an interactive client experience, and automated optimization — not by out-OCR-ing a corpus moat.
Mixed inputs converge to one schema; a single pure engine feeds a fan-out of analytics that each call it hundreds-to-thousands of times; the results converge into one response surface, rendered once into three deliverables for two kinds of user.
Parity on the foundation; the wedge is automated optimization, multi-variable surfaces, breakpoint enumeration, adjustable assumptions, and a genuinely interactive client surface.
| Capability | Lumio | Holistiplan | FP Alpha | RightCapital |
|---|---|---|---|---|
| Ingest & data | ||||
| OCR ingest — 1040 / IRS transcript | ◐ | ● | ● | ○ |
| Tax-prep import (ProConnect / Drake / Lacerte) | ◐ | ◐ | ○ | ◐ |
| Tax engine | ||||
| Deterministic, tax-year-versioned engine | ● | ● | ◐ | ◐ |
| Line-level explainable trace (every figure sourced) | ● | ◐ | ○ | ○ |
| Scenario & optimization | ||||
| Side-by-side scenarios (Roth / gains / charitable) | ● | ● | ◐ | ● |
| Automated constrained optimization ("highest under constraint") | ● | ○ | ○ | ○ |
| Multi-variable optimization (2-D surface / heatmap) | ● | ○ | ○ | ○ |
| Breakpoint / cliff enumeration to the dollar | ● | ◐ | ○ | ○ |
| Multi-year lifetime projection | ● | ● | ◐ | ● |
| Explicit advisor-adjustable assumptions (discount / growth / inflation) | ● | ○ | ○ | ◐ |
| Client experience | ||||
| Client-facing interactive explorer (live sliders) | ● | ○ | ○ | ◐ |
| White-labeled PDF report | ● | ● | ● | ● |
| Hosted, access-controlled interactive report | ● | ◐ | ● | ● |
| Practice & trust | ||||
| Practice OS (roster · status · docs · comms · §7216 · IRS notices) | ● | ○ | ○ | ◐ |
| Estate / insurance breadth | ○ | ● | ● | ◐ |
| SOC 2 / RBAC / per-tenant encryption | ◐ | ● | ● | ● |
| Pricing model | ||||
| Go-to-market pricing | Hybrid — advisor seat + metered reports (resale-friendly); enters under the incumbent's raised price | Per-seat advisor subscription (recently increased) | Per-case (~$13–15 / case) | Per-seat advisor subscription |
Competitor capabilities reflect Lumio's own competitive analysis (consolidated-plan.md §0) and
public marketing as of early 2026; marketing routinely outruns shipped depth, so these are positioning
estimates pending direct validation (plan §13). Lumio's marks reflect the designed product per the
consolidated plan; the current build is a structural demo on placeholder tax rules — see Build status below.
The federal 2025 rule set is verified and pinned by tests — constants and worksheet mechanics for the areas 17 years of practice flag as high-risk (Social Security taxability, CTC/ACTC) are asserted by the suite, not assumed. Hawaii N-11 values are not yet verified and remain gated. CI still fails any production build pointed at a placeholder rules file.
Pure-function engine with a 14-step traced pipeline; the candidate-enumeration + adaptive-bisection optimizer matched to a brute-force sweep within $1; 230+ Python tests, 28 Vitest tests, clean build.
Hawaii N-11 verified values; SOC 2 + RBAC + per-tenant KMS before any multi-tenant data; a real OCR vendor backend.
"Lumio" is a working name pending a formal trademark-clearance search in software / financial-services classes. Irreversible brand buildout is on hold until clearance.
Operating instructions here live in hand-written prompt files rather than a
CLAUDE.md, because I was building the same thing from scratch against different
providers to compare them; the CLAUDE.md stays a general guide, not plan-specific direction.
The tax constants are pinned by tests, not because a model got one wrong, but because 17
years of returns told me which areas carry the risk, and those are the ones I had hardened
first. Verification closes outside the agent: I check the numbers myself and against returns
run independently in software I trust, with skills that populate that software from here, so
an error surfaces on either side of the input tree. The client deliverable took the most back
and forth, since the reader is usually a layperson; the rest was hardened by an adversarial QA
loop of three iterations over 27 fictional client profiles, implemented as Codex work orders
and reviewed, plan and every diff, before anything landed.