Review before use
Values below 85% confidence, lines that mix cost roles and low-confidence OCR are held for review. Only approved values feed a model.
- Home
- Platform
Platform
The capital committee above your systems of record.
Capibud does not replace Primavera, SAP or your planning tool. It reads the files those systems and your teams already produce, runs every decision on the same correlated scenarios, routes the result through your delegation of authority and records what was approved.
Available to design partners today as a deployment inside your own environment, on hardware or cloud you control. Managed regional hosting and certifications are on the roadmap; the security section separates the two.
4,000Correlated scenarios behind every P50, P80 and P90.
16Shared factors, so a common shock hits every exposed project in the same scenario.
37Researched industry packs, versioned and pinned to each decision.
~0.8 sFor a seven-project book with 4,000 scenarios and its optimiser.
Where Capibud sits
Above the systems you already run
Your schedules stay in your scheduling tool and your ledgers stay in your ERP. Capibud does not hold your system of record. It reads what those systems export, and what your teams write around them, and turns it into a decision the committee can size, sign and replay.
Your systems of record
- Schedules from Primavera P6 or Microsoft Project
- Cost ledgers and budgets from SAP or another ERP
- Planning workbooks with no common template
- Steering memos, board decks and scanned letters
Capibud reads and simulates
- Files read as they are, every value traced to its source
- Low-confidence values held for a person to review
- A researched industry pack as the baseline
- Every decision on the same correlated scenarios
The committee decides
- The book sized against the bad year
- Routed through your delegation of authority
- Recorded so it can be replayed
- Watched by the brain once it is approved
Capibud works with what P6, MS Project and your ERP export today (P6 XER, P6 and MS Project XML, Excel and CSV); native connectors are rolling out.
The layers
Seven layers, one book
Each layer does one job and hands a traceable result to the next. Together they take a pile of files to an approved, recorded decision that stays watched.
Layer 01Available today
Read files, with provenance
Capibud reads un-templated workbooks, CSV ledgers, Word memos, PowerPoint decks, PDFs and scanned documents. Each value keeps its source: the sheet and cell, the slide and shape, or the page, line and position on the scan.
Values below 85% confidence, lines that mix cost roles and low-confidence OCR wait for a person to review them. Only approved values feed a model.
Local OCR85% review threshold15 MB per file
What Capibud reads → Layer 02Available today
Start from a researched baseline
37 industry packs, from refining and LNG to data centres and transmission, supply sourced parameters, price feeds, risk archetypes and project templates. A new project does not start from a blank sheet or from someone's memory of the last one.
Packs are versioned and pinned to each decision, so the baseline a committee relied on is the baseline the record replays.
37 packsVersioned and pinned
Browse the industry briefs → Layer 03Available today
Simulate on shared scenarios
Schedule, cost and value are simulated together across up to 16 shared factors. Projects exposed to the same commodity, supplier or rate move together, so one transformer supplier or one EPC crew shows up as a concentration rather than as several independent risks.
Every option gets a P50, P80 and P90 range for cost, schedule and value, instead of one number with no distribution behind it.
4,000 scenarios16 shared factors~0.8 s seven-project book
Layer 04Available today
Size against the bad year
The capital committee chooses a risk appetite and solves the whole book, across all four decisions, for the highest expected value within it. The bad year is the average outcome in the worst 5% of scenarios.
Alternatives, stress tests and reverse stress tests (the smallest shock that breaks the plan) are shown beside the recommended plan.
The capital committee →Layer 05Available today
Decide through your delegation of authority
Every change is previewed before it is applied. Requests are routed by your delegation-of-authority matrix, with maker-checker, escalation and cumulative-change rules. The final approval applies the change, which can be undone.
Roles (viewer, analyst, approver and admin) are read from the database on every request, never from the request itself.
Preview before applyMaker-checkerUndo
Layer 06Available today
Keep a replayable record
The audit log is append-only and hash-chained. Each engine run stores digests of its inputs, parameters and output, its seed and engine version, and can be replayed to an identical result. Committee decisions add votes, dissent and tripwires.
The decision record →Layer 07Partial
Watch the book: the brain
The monitoring layer runs on a schedule across every workspace. A market move, filing or actual arrives; Capibud finds the exposed projects and holdings, prices the effect in money, drafts a proposal, verifies it in a sandbox and sends it to the right person's queue. Nothing changes the plan until a person approves it.
Outcome learning needs months of real decisions, which is why this layer is marked partial.
The brain →What Capibud reads
The files your teams already produce, as they are
Capital decisions are made on spreadsheets that follow no template, memos with tracked changes, steering decks and scanned letters. Capibud reads them without re-keying and shows where every value came from.
| Format | What Capibud reads | Status |
|---|---|---|
| Excel and CSV | Several tables per sheet, merged multi-row headers, notes and units in titles, transposed tables, period columns such as FY25, Q3 FY25 and 2026E, ₹ crore and lakh, $ million and £ million | Available today |
| Word and PowerPoint | Body text and tables, tracked changes (sent to review), slide tables, embedded chart data and speaker notes | Available today |
| PDF and scans | Text PDFs and ruled tables; scanned pages and images through local OCR with rotation, deskew and per-word confidence. A deliberately degraded test scan reads at 94–96% mean word confidence with every table cell correct | Available today |
| SEC filings | 10-K, 10-Q and 8-K filings and XBRL company facts, fetched on a schedule, compared with the previous filing and passed to the brain when a figure changes | Available today |
| Native connectors | Live connections to P6, MS Project and ERP systems. Today Capibud works with what they export: P6 XER, P6 and MS Project XML, Excel and CSV | Roadmap |
Mapping without sharing values
When a column layout is unfamiliar, an optional language model suggests the mapping. It sees column headers and a few row labels, never cell values, and its reply is validated before use.
Current limits
15 MB per file and 10 OCR pages per document. No handwriting. OCR runs locally, so documents do not leave the server.
Three kinds of model
The engines produce every number
The simulation and optimisation engines produce the ranges, the sizing and the stress tests. Around them, three kinds of model each do a narrow job, and only one of them is a language model, which is optional.
Rules
Versioned rules
Thresholds that catch known patterns quickly, for example a commodity moving 1.5 standard deviations or a contractor's listed peer falling 25%. Each rule has a version, so a proposal can say which one fired.
Fitted
Fitted models
Reference-class overrun priors, signal-to-probability calibration and factor correlations, fitted and then frozen. The signal-persistence model has an out-of-sample Brier score of 0.172, where a coin flip scores 0.25.
Optional
Language models, number-checked
Read unstructured text, suggest mappings and draft explanations. Every number in AI-written text must match the recorded run exactly, or the text is rejected and the deterministic version is used. Off unless configured.
Security and deployment
Your environment, your controls
Capibud runs inside an environment you control. Without explicit configuration it makes no external network calls, and a test confirms it. The controls below are available today; the table after them separates today from the roadmap.
Append-only, hash-chained log
Every change is written to an append-only log that is hash-chained and can be verified. Signed checkpoints can be checked with a public key.
OCR runs locally
Scanned documents are read on your server. They are not sent to an outside service to be recognised.
AI layer off unless configured
Language models are switched off by default. Capibud works fully without them, and makes no external calls unless you configure them.
Models never see cell values
When a language model helps map an unfamiliar file, it sees column headers and a few row labels only. Its reply is validated before use.
Kill switch
Operators can stop any automated job. A held job completes without changing anything.
Maker-checker
Changes are previewed and routed by your delegation of authority. Approvals are re-validated against what the approver actually reviewed.
Undo
Every applied change can be undone, and the undo is recorded like any other change.
Tenant isolation and roles
Server-side, revocable sessions; deny-by-default access between tenants; roles read from the database on every request.
No funds moved
Capibud does not move funds or place orders. "Apply" changes a plan inside Capibud and nothing else.
Today and roadmap Production boundaries
| Area | Available today | Roadmap |
|---|---|---|
| Deployment | Runs on hardware or cloud you control, as a single node or container | Managed service in India, the EU and the US; sovereign and air-gapped pods with customer-held keys |
| Network and AI | The AI layer is off unless explicitly configured. Without configuration, a test confirms Capibud makes no external network calls. OCR runs locally | In-country open-weight models for India-only tenants; EU-pinned model endpoints |
| Identity and access | Server-side, revocable sessions; tenant isolation with deny-by-default access; roles (viewer, analyst, approver, admin) read from the database on every request; OpenID Connect single sign-on, tested against a local identity provider; rate limits | SAML through a direct integration, SCIM provisioning, testing against live Entra ID and Okta tenants |
| Approvals | Preview before every change; maker-checker; approvals re-validated against what the approver reviewed | No change planned |
| Audit | Append-only, hash-chained log with verification; signed checkpoints that can be verified with a public key; run replay to an identical result; backup and restore drill | Independent custody of checkpoints; RFC 3161 timestamping |
| Certifications | None yet | SOC 2 Type II and ISO 27001 programme; CERT-In-empanelled penetration test in India |
| Data feeds | Public sources (FRED, SEC EDGAR, ECB and others) for demonstration | Licensed market data before any customer use |
| Interface | Nine languages, machine-assisted translations; commands by text or voice; any display currency at plan or ECB rates | Arabic; professionally reviewed translations |
Current limits
What Capibud does not do yet
A platform that asks a committee to trust its ranges should be plain about its own boundaries. These are the ones that matter most when you plan a first decision with us.
- File exports, not live connectors. Capibud works with what P6, MS Project and your ERP export today (P6 XER, P6 and MS Project XML, Excel and CSV); live connectors are rolling out.
- File limits. 15 MB per file and 10 OCR pages per document. Handwriting is not read.
- No certifications yet. A SOC 2 Type II and ISO 27001 programme is on the roadmap. Until then, Capibud is deployed inside your environment, under your controls.
- No managed hosting yet. Managed regional hosting is on the roadmap; today Capibud runs on hardware or cloud you control.
- Single sign-on. OpenID Connect has been tested against a local identity provider only. Direct SAML, SCIM and live Entra ID and Okta testing are on the roadmap.
- Demonstration data feeds. Market data comes from public sources for demonstration. Licensed market data is required before any customer use.
- No customer back-test yet. Nothing has yet been back-tested on a customer's own history. The first Decision Sprints will provide that evidence.
- The brain is partial. Outcome learning needs months of real decisions.
- Translations. The eight non-English interface languages are machine-assisted, not professionally reviewed.
Questions buyers ask
Before the first workshop
The decision itself is covered on the capital committee page, and the audit trail on the decision record page.
Does Capibud replace Primavera, SAP or our planning tool?
No. Your schedules stay in your scheduling tool and your ledgers stay in your ERP. Capibud reads what those systems export today (P6 XER, P6 and MS Project XML, Excel and CSV) and what your teams write around them; native connectors are rolling out.
What files can Capibud read?
Un-templated Excel and CSV workbooks, Word memos with tracked changes, PowerPoint decks, PDFs and scanned documents through local OCR. Every value keeps its source, and values below 85% confidence wait for a person to review them before they feed a model.
How does Capibud decide where the budget goes?
Every option is simulated on the same 4,000 correlated scenarios across up to 16 shared factors. The capital committee sets how much of the bad year it will accept, Capibud solves the whole book for the highest expected value within it, and the result is routed through your delegation of authority.
Where does Capibud run?
Today it is available to design partners as a deployment inside your own environment, on hardware or cloud you control. Managed regional hosting and certifications are on the roadmap, and the security section separates the two.
See the platform on your own files.
A Decision Sprint runs one real decision on your files in six weeks: up to three projects or one portfolio, one industry pack and two data sources.