[ DEBT_CAPITAL_MARKETS ]

Contract Q&A and Covenant Tracker

2026Document IntelligenceIn-depth

What it is

Ask the same twenty questions of every contract that comes in, and track the obligations you find.

Client work is shown without identifying imagery

[ THE_PROBLEM ]

Why this existed

Somebody has to read every incoming contract looking for the same twenty things. It is expensive, slow, and the one time it is done carelessly is the time it matters.

[ WHAT_WE_BUILT ]

What we built

Upload a PDF and ask ad-hoc questions, or — the real workflow — build a reusable question preset once and run it against every new document. Each question is tagged with how it should be answered: fast vector retrieval for questions with a findable answer, or a chunk-by-chunk full-document scan for questions where a miss is expensive and you need coverage rather than speed. Mixed presets run both paths in one job. Every run is stored with per-question status and answers, exportable to CSV, plus a dedicated covenant tracker that extracts and monitors the obligations found.

Ad-hoc question answering against an uploaded PDF or text document
Reusable question presets run against every new document
Per-question mode: fast vector retrieval, or chunk-by-chunk full-document scan
Mixed presets running both paths in a single job
Run history with per-question status and stored answers
CSV export of results
Covenant tracker extracting and monitoring obligations found

[ HOW_IT_IS_USED ]

How a company uses it

A trustee, lender or legal ops team replaces "a paralegal reads all 60 pages looking for the negative covenants" with a preset that runs in minutes and produces the same table every time — with a full-scan mode for the clauses you cannot afford to miss.

Built with

ConvexReact 19OpenRouterpdf-parseVector search

[ COMMON_QUESTIONS ]

Questions clients ask

Why have two different answering modes?

Because they fail differently. Vector retrieval is fast and cheap but can miss a clause that is worded unusually. A full-document scan reads every chunk, so it is slower and costs more but gives you coverage. For "what is the interest rate" retrieval is fine; for "list every negative covenant" you want the scan. Letting the user tag each question is the honest solution.

Can it replace a paralegal review?

It replaces the first pass, not the judgement. What it reliably does is produce the same structured table from every document in minutes, with the source available for each answer. A human still confirms anything consequential — but they are checking rather than hunting.

What document types does it handle?

PDF and text out of the box. The pattern extends to any contract family where you ask a repeatable set of questions — loan agreements, leases, supplier contracts, insurance policies. The preset is the reusable asset.

[ RELATED_WORK ]