All Data

Permission-aware chunking for enterprise RAG — ACL metadata on every chunk, so retrieval never leaks across access boundaries.

Score the retrieval step of your RAG pipeline — Recall@k, Precision@k, MRR, nDCG — and get a ship / hold / fix verdict with every miss named. Python + TypeScript, any vector DB, gates retrieval in CI. Pairs with AutoChunk: chunk there, grade here.

Master prompt + 8 focused prompts + 24-question bank across Sales/Marketing/Finance/Ops. Five-rule honesty layer keeps the output defensible. For any AI tool that runs code on uploads.

12 ICP personas + bias-check protocol. Pressure-test positioning in a weekend.

Grade whether an exported CSV is structurally fit to be processed before any row-level tool touches it. Per-column CLEAN / DIRTY / MISSING plus one file verdict — READY / FIX / BLOCK — with a hard structural gate: a ragged row, mangled header, or missing required column blocks the file no matter how clean the cells that parsed look (the sample is all-CLEAN, quality 100, and still BLOCKs). Runnable Python engine + workbook. The upstream gate for the whole Document Processing line; grades the file, not people.

Reconcile 3+ sources on one key — a bank export, a processor export, and your ledger; or three feeds that must agree. Classifies every key AGREED / DISPUTED / INCOMPLETE (separating a value conflict from a coverage gap, which a two-list tool can't) and rolls up to RECONCILED / FIX / NOT RECONCILED, with a one-disagreement gate: a single disputed key blocks the reconciled verdict however many keys agree. Runnable Python engine + workbook. Deterministic, offline; reconciles data values, not people.

A column-aware PII readiness check for CSV / CRM exports. Classifies each column by the PII it holds and whether it's handled to the minimum its class requires — HANDLED / MASK FIRST / EXPOSED per column, SAFE TO USE / REMEDIATE / HOLD for the dataset — with a regulated-exposed gate: one open regulated column (an SSN with no handling) holds the whole dataset, 83% handled notwithstanding. Column-grain, not a cell-by-cell text scan. Runnable Python engine + workbook. Deterministic, offline; flags columns, not people. Not legal advice.

Audit a spreadsheet's formula layer — not its numbers — for the faults that silently corrupt a model: error cells, broken links, hardcoded magic numbers, copy-paste inconsistencies, and volatile dependencies. Reads a real .xlsx and returns SOUND / REVIEW / BROKEN per formula cell and a model gate where one error cell breaks the whole model — the shipped sample scores 92 and is still BROKEN because one cell carries a #REF!. Runnable Python auditor + workbook. Deterministic, offline; audits formulas, not people. A clean total is not a sound model.

Screen a bank or card feed before it reaches the ledger. Auto-categorize by your own keyword rules and run four anomaly checks → CLEAN / REVIEW / FLAG per row and a batch POST READY / REVIEW ROWS / HOLD, where one uncategorized row, bad date, bad amount, or duplicate holds the whole batch. No AI, deterministic. One spreadsheet.

The deterministic layer that judges extracted or keyed document fields against fixed rules — format, range, required, allowed-value, and cross-field math (subtotal + tax = total) — and returns PASS / REVIEW / REJECT per record. A runnable Python engine + a workbook reproduce the same rules, with a hard-fail gate that rejects any record with a critical failure (a missing total, a math error) however high it scores — a record can't be 90% valid. Type-agnostic: swap the rule pack for POs, applications, claims, intake forms. The validation step the AI Document Extraction Kit hands off to. Validates your own documents, not people. Not legal advice.

The deterministic pre-share safety gate for documents. A pattern pack of fixed detectors scans a document's text for PII across three sensitivity tiers — SSNs, cards, bank routing, health markers, dates of birth, addresses, emails, phones, IPs — and returns CLEAR / REDACT FIRST / DO NOT SHARE per document, with a library rollup. An exposure gate blocks any document with an unhandled high-sensitivity finding however high its coverage — one open SSN is dispositive. It flags what to redact and what's still exposed; it does NOT redact, is not a certified tool, and does not score or rank people. Tunable patterns and tiers. Validates your own documents; not legal advice. You remain responsible for what you share.

Turn invoices, receipts, and forms into structured data with AI that flags low-confidence fields for review instead of guessing. A runnable extractor (zero deps, runs keyless in demo, exits non-zero on review), editable JSON schemas, a QA tracker, and a playbook. Flag, don't guess — no accuracy guarantee.