Available now
In build
The whole team
Nineteen specialists, each with a defined job and an honest status label.
See all nineteenBottom-up TAM built from your own inputs, a cash projection with a receivables lag, sensitivity ranked with the method shown, seven canvases filled cell by cell. A claim with no source is not a value this code can construct — a top-down market total with no publisher, URL and retrieval date is refused before it is ever stored.
What it does
Every card's status label is derived from the backend's own capability module.
Every item is a claim row with a mandatory source kind, so a SWOT cannot hold a figure that does not say where it came from. An LLM-assisted first draft needs a model key; hand-entered analysis does not.
Works todayThe four crossings are generated as questions the owner answers. An action names the claims it was derived from, so a reader can walk a recommendation back to its evidence.
Works todayTAM/SAM/SOM computed from the owner's own inputs, with the working readable back. A top-down figure with no publisher, URL and retrieval date is refused, not stored as a rounder-sounding cross-check.
Works todayRevenue earned is not counted as cash received — a days-sales-outstanding lag delays it. Break-even returns as a revenue level with the month it is reached. Sensitivity ranks the assumptions by effect on closing cash and reports the perturbation it used, so the ranking is reproducible rather than asserted.
Works todayBusiness model canvas, go-to-market, buyer persona, OKRs, PESTEL, Porter's Five Forces and Lean Canvas, filled from a plain-language interview rather than handed over blank. A cell nobody answered stays empty.
Works todayA cadenced sweep of what rests on nothing checkable, sent through Wavy's own consent-gated rail — no second sender, no opt-in means no send. Now aggregates competitor, pricing and opportunity sections alongside the freshness sweep, and persists what it composed so a past digest reads back as real history, not just a last-result string. Stays silent and unsent when there's nothing to report.
Works todayPESTEL and Porter's Five Forces are canvases the owner fills from what they have personally seen, each cell a sourced claim. Generating the factors needs a grounded retrieval source this deployment does not have.
Built, with limitsNot an ordering column — a reordered plan stays correct. The conflicts route reports circular and unschedulable dependencies instead of quietly emitting a broken plan.
Works todayDraws only from this workspace's own saved analyses and refuses to answer from anything else. Conversations and their citations now persist, so a past answer reads back with the claims it cited. Keyword-ranked citations, the regulated-advice gate and the stored-claims fallback all run with no credential; generating the answer itself needs a model key.
Needs a keyFreshness state and the estimate-versus-published-source split across every stored artifact, plus re-verification. Version snapshots now exist too — each one a deliberate action carrying a required change reason, not an automatic save on every edit — with a field-level diff between two versions, so a reformatted figure doesn't read as a changed one.
Works todayAn owner-registered URL is fetched through the SSRF guard and that domain's own robots.txt, never captured when disallowed. Two real captures are diffed mechanically — never invented — and categorised by keyword heuristic; an LLM can optionally narrate a diff, but every change already has a real mechanical summary with no credential.
Works todayEvery claim is compiled straight from a competitor's own tracked CompetitorChangeEvent rows — source URL and retrieval date always present — so a competitor with nothing tracked yields an empty card, never a fabricated one.
Works todaySections are structurally assembled from the owner's own stored SWOT, market-size and forecast claims — every bullet is a claim's own text, never a rewrite, so the core capability needs no credential. A rule-based check flags a narrative growth claim against a flat or declining forecast rather than letting a polished draft hide it.
Works todayWatch items and the benchmark computation work fully against your own database. Every observation is logged by a person who checked, with price, currency, listed unit and source URL — never an automated scrape parsed into a number. Comparisons never cross currencies.
Works todayPattern-scan over what's already in this workspace — unreviewed competitor pricing changes, price-benchmark extremes, unaddressed SWOT weaknesses and threats. Every signal names the stored rows it rests on and carries a confidence level set by how many corroborate it, never a bare score, and a repeat scan never duplicates a signal already on file.
Works todayCategory-level suggestions are real and need nothing configured: derived from your own SWOT weakness and threat claims against a fixed rule table, each citing the claim that prompted it. Named candidates are built too, against OpenStreetMap — free, keyless, and every name carries the directory URL it came from, never one a model invented. It is off until an operator switches it on, because it makes this server query a third party on your behalf.
Needs a keyTime-limited (14 days by default, up to 90) and revocable links across every Yukti artifact, through one shared mechanism. A comment-permission link lets someone comment with no platform account of their own; neither a comment link nor a view-only one ever gains write access to the underlying artifact.
Works todayRenders and stores a complete, self-contained HTML document stamped with its generation date. No PDF or slide-deck renderer is installed on this deployment, so that path returns an honest failed export naming the gap rather than a silent one.
Works todayEvery figure carries one of these
A claim that does not say where it came from is not a value this code can construct. A top-down market figure with no publisher, URL and retrieval date is refused before it is stored — the tool defaults to "insufficient data" rather than a plausible-sounding number.
| Source kind | Means | Required to record it |
|---|---|---|
| owner_input | Stated as fact about their own business ("we open at 9"). | A line saying what it rests on. Shown as "You told us: …" |
| owner_estimate | The owner's own guess about the future. | Labelled as a guess everywhere it is shown, never blended with a sourced figure. |
| owner_record | Read off the owner's own records — books, POS, invoices, bank statement. | The only kind accepted when recording what actually happened in a forecast month. |
| workspace_data | Computed from data this deployment already holds — contacts, leads, sends. | Nothing extra — it is counted, not asserted, and is independently checkable. |
| published_source | An external document — a report, a filing, a news article. | A URL, a publisher, and a retrieval date. Missing any one of the three is refused. |
| yukti_inference | Yukti's own reasoning over other stored figures. | Must name what it was derived from — an untraceable inference is refused. |
How it looks
Limits
Every figure here is a constant in the codebase.
| Cash projection horizon | 1–60 months |
|---|---|
| Sensitivity perturbation | 10% default, up to 100% |
| Strategy digest cadence | 1–90 days, default 7 |
| Digest — items shown per section | 6 |
| TOWS pairs returned by default | 40 |
What it refuses to do
Before you switch
Tell us the plan you are building and we will show you exactly what Yukti carries — and the eight things it still cannot source.
Questions
It can build a bottom-up estimate from numbers you supply — how many customers are in reach, what share you could serve, what they would pay — and show the working. A claim that does not say where it came from is not a value this code can construct. A top-down market figure with no publisher, URL and retrieval date is refused before it is stored — the tool defaults to "insufficient data" rather than a plausible-sounding number.
Three kinds that are never allowed to blend: what you told Yukti as fact, what you yourself guessed, what a published document says with its publisher and retrieval date attached, and what Yukti itself worked out and can name the inputs for. Every figure on every artifact is labelled with one of these, every time it is shown.
Revenue earned is not treated as cash in hand. A days-sales-outstanding lag delays when each month's revenue actually turns into cash, which is the correction that stops an owner on 60-day payment terms from being told they can make payroll in a month they cannot.
Yes, one month at a time, but only from your own records — the till, the bank statement, the invoice book. A remembered figure set beside a projection would produce a variance made of two guesses that reads as a measurement, so Yukti refuses anything that is not an owner_record. The assumptions themselves are never rewritten to match.
Competitor tracking watches URLs you register — fetched through an SSRF guard and that domain's own robots.txt — and diffs two real captures mechanically, never by invention. Pricing benchmarks are built from observations a person logged, never an automated scrape parsed into a number. Naming an actual partnership candidate is the one piece still not built: no business directory or maps API is connected, so that route returns an honest 'blocked, no source' rather than a name a model invented.
No. Its generated prose is checked against the same regulated-domain gates the rest of the product uses, and a match returns a redirect to a qualified professional instead of the text.
