Case study · Data and analytics
Pairenthesis
Verifiable, plain-language analysis for data that has to be defended
- Client
- Toor Technologies product
- Industry
- Data and analytics
- Category
- Website Development
- Services
- Website Development, Cloud Services, Data Science
- Website
- pairenthesis.com
The Challenge
Most teams that work with spreadsheets and exports spend more time cleaning data than analyzing it, and the people with the questions are rarely the people who can write the queries. Chat-style AI tools promised to close that gap, but in regulated settings such as government, finance, healthcare, and infrastructure, a number that a model guessed at is a liability, not an insight.
The challenge was to keep the convenience of asking questions in plain English while removing guesswork from the answers. Every figure had to be computed in code, every transformation had to be reviewable before it took effect, and the whole trail had to be auditable later. Data residency and a bilingual interface were requirements from the start, not add-ons.
What We Built
Pairenthesis is a governed data operations platform built on Next.js and React, with Supabase for data and authentication, deployed on Vercel. Users connect files and systems, and a deterministic engine profiles every column for types, nulls, outliers, and format violations before any AI model is involved.
When a user asks a question in plain language, the platform turns it into a parameterized query, shows that query alongside the result, and returns a verified figure with a one-line explanation. Inbound records pass through a gate that checks schema validity, statistical outliers, and clearance-level redaction. Flagged items are held for a person to approve, and nothing commits automatically.
Forecasts are computed deterministically as scenario-based projections with confidence intervals, and a daily brief summarizes data health, open issues, and improvements. Every computation, answer, and change is written to an append-only, hash-linked audit ledger. Data stays in the customer's region, and the interface is available in English and French.
Key Features
Deterministic profiling engine
Profiles every column for types, nulls, outliers, and format violations before a model runs, so cleaning starts from facts about the data.
Plain-language questions, verified answers
Ask in English or French. Each answer shows the figure, the parameterized query behind it, and a one-line explanation.
Inbound review gate
Checks each record for schema validity, statistical outliers, and clearance-level redaction. Findings wait for a person to authorize them.
Deterministic forecasting
Scenario-based projections with P10 to P90 confidence bands, computed in code rather than estimated by a model.
Append-only audit ledger
Every computation, answer, and change is recorded in a hash-linked ledger that cannot be edited after the fact.
Daily brief and data residency
A morning summary of data health and open issues, with data stored and processed in the customer's own region.
Outcomes
- Every figure shown to a user is computed in code and traceable to the query that produced it.
- Nothing changes the data without a person approving it, which keeps the platform usable in regulated environments.
- A complete, tamper-evident record of analysis is available for audits without extra work from the team.
- The interface works in both English and French, and data stays in the customer's region.
Technologies
- Next.js
- React
- Tailwind CSS
- Supabase
- Vercel
