Built for Court Reporters
CasePrepd isn't a generic AI tool with a court-reporting label slapped on it. The logic inside was built by a working court reporter to handle the things that actually go wrong when you point an AI at a deposition transcript. Here's what that looks like in practice.
Offline and Private
CasePrepd makes no network calls. After installation, the app runs entirely on your computer using AI models that ship with the installer. There's no login, no API key, no cloud component to opt out of. Disconnecting from the internet has no effect on how the app behaves.
This matters for court reporters in two specific situations:
- No reliable internet — courthouse wifi is frequently blocked or unavailable. Hotel, train, plane, and rural-deposition connectivity is unpredictable. Cloud AI tools fail in all of these settings; CasePrepd doesn't.
- Material that can't be uploaded — sealed cases, grand jury transcripts, HIPAA-protected medical records, attorney-client privileged exhibits, and any documents covered by firm or court policies prohibiting use of cloud AI. Sending these through ChatGPT or comparable services is, at minimum, a confidentiality issue.
Because processing happens locally, none of these constraints apply. What you do with the documents is between you and your computer.
Why "Generic AI" Falls Short on Transcripts
If you've tried feeding a deposition into ChatGPT or a generic NLP tool, you've probably seen these problems:
- The same witness shows up under five spellings (one for each OCR error) and you can't tell which is right.
- Your vocabulary list is half "Q", "A", page numbers, and "DIRECT EXAMINATION".
- The tool thinks "Di Leo: Objection" and "DI LEO 1 Q" are two different people.
- Names get listed alphabetically with no sense of who's actually a key witness vs. who's a passing mention.
- When the tool decides to "merge" similar terms, there's no way to see what it merged or undo a mistake.
CasePrepd was designed around fixing each of these specifically. The features below are the result.
The Name Canonicalizer
OCR'd transcripts produce name variants. Lots of them. A single witness might appear ten different ways across a deposition — sometimes with letters swapped, sometimes with extra punctuation, sometimes truncated by a scanning artifact. CasePrepd collapses these into one entry and picks the most likely correct spelling.
Example
A generic vocabulary extractor on a single transcript:
- Arthur Jenkins (31 occurrences)
- Anhur Jenkins (4 occurrences)
- Arthur Jenidns (2 occurrences)
- Arlhur Jenkins (1 occurrence)
CasePrepd:
- Arthur Jenkins (38 occurrences)
The merged variants and the reason each was rejected are visible by right-clicking the row — see "You Can See What Got Merged" below.
How It Picks the Right Spelling
The canonicalizer doesn't just pick the most frequent variant — that would be wrong when one variant is in the dictionary and the others aren't. It uses a tiered strategy:
- If exactly one variant is a known name (dictionary or name dataset), that variant wins. "Arthur" is in the dictionary; "Anhur" isn't.
- If none of the variants are known (rare or exotic name), a weighted score decides — combining document confidence (OCR quality), frequency, and an OCR artifact penalty.
- If multiple variants are known ("Smith" vs "Smyth", both real names), the weighted score breaks the tie based on which appears more confidently in your documents.
Transcript Artifact Stripping
Court transcripts have format conventions that generic NLP doesn't understand. The name "Di Leo" might appear in a transcript as any of:
DI LEO: Objection
DI LEO 17
Mr. Di Leo
Di Leo:
A generic vocabulary tool gives you all five as separate entries. CasePrepd recognizes the patterns — line numbers attached to speaker labels, Q/A notation, colon-suffix speech attribution, and common title prefixes ("Mr.", "Mrs.", "Dr.") — and folds them into one entry for Di Leo. Variants like "THE WITNESS DI LEO" that don't match a pattern are caught separately by the fuzzy-match canonicalizer above.
Transcript Section Awareness
Depositions and trial transcripts have standard section headings: DIRECT EXAMINATION, CROSS-EXAMINATION, REDIRECT, RECROSS, and so on. A naive vocabulary extractor latches onto these because they're high-frequency and look important — you end up with vocabulary rows like "Di Leo Direct" or "Smith Cross-Examination" that aren't really terms at all.
CasePrepd knows the standard section keywords and filters out spurious combinations of names + section words before they ever reach your vocabulary list.
Automatic Page Cleanup
Stenotype-formatted transcripts have features that need to be stripped before any NLP can work on them. CasePrepd handles these automatically in its preprocessing pipeline:
- Remove line numbers — the 1–25 margin line numbers on every transcript page are detected and removed. Without this, every page produces 25 spurious "terms" that are just integers.
- Remove headers/footers — repeating case caption, deposition title, and page number lines. Detected by frequency analysis (text that repeats on most pages), so it works on transcripts from any reporting firm.
- Remove index pages — the alphabetical concordance/index pages at the end of transcripts are stripped so the index terms don't dominate the vocabulary list.
- Clean page boundary artifacts — when PDF extraction collapses page breaks, line numbers and reporter initials can merge into body text. This cleaner detects and removes them.
- Clean transcript artifacts — standalone page numbers and inline page:line citations embedded in some transcript formats are scrubbed.
Each of these is a separate checkbox in Settings → Text Preprocessing, on by default, and can be turned off individually if you're working with documents where one misfires (e.g., a non-transcript exhibit).
Role Detection
For every person CasePrepd extracts, it looks at the language around each mention to figure out their role in the case. The current role tags are:
- Plaintiff and Defendant — matched on context like "plaintiff Smith" or "defendant Jones"
- Plaintiff attorney and Defendant attorney — matched on "plaintiff's counsel" / "defendant's lawyer" patterns
- Treating physician — matched on the explicit "treating physician" / "treating doctor" phrasing
- Medical professional — "Dr. X" titles and "X, a nurse/physician/surgeon" phrasing
- Witness — "witness X", "X testified", "X was called to testify"
The role appears in the Role/Relevance column of the Vocabulary table. When you're scanning a list of 80 people from a multi-witness case, knowing that "Dr. Martinez" is a treating physician saves real time during prep. The role-tag list is intentionally narrow — only patterns that fire reliably on real transcripts make it in.
You Can See What Got Merged
Most AI tools are black boxes. They merge similar terms, and you have no way to confirm they merged the right things — or notice when they merged two different people who happened to share a similar spelling.
CasePrepd shows its work. Any name row that resulted from a merge is marked, and right-clicking it opens the Alternatives dialog. You see every variant that was rolled into the canonical entry, each variant's count, and a human-readable reason for why it was accepted or rejected. Examples:
Arthur Jenkins — canonical
Picked because "Arthur" appears in the name dictionary; the others don't.
Rejected variants:
Anhur Jenkins— "Anhur" not a known name; likely OCR artifact (h→hn)Arthur Jenidns— "Jenidns" not a known surname; likely OCR artifact (k→id)Arlhur Jenkins— "Arlhur" not a known name; likely OCR artifact (t→l)
If the canonicalizer ever picks the wrong winner — or merges two genuinely different people — you can see exactly what happened and override it with a Skip rating. The model learns from that, so the same mistake gets less likely on future jobs.
The Things You'd Expect from a Generic Tool, Done Better
Beyond the court-reporting-specific features, the everyday capabilities are tuned for legal work too:
- Medical terminology — a specialized model recognizes drugs, diseases, anatomy terms, and medical procedures. Med-mal and personal-injury cases surface the medical vocabulary that generic tools miss.
- Multi-document awareness — a name appearing in 6 of 8 transcripts is flagged as more important than one appearing in just one. Helps you build a unified job dictionary across multi-deposition cases.
- Per-user learning — the Keep/Skip thumbs train a model on your preferences. After a few jobs, it gets noticeably better at flagging what you personally care about — not what some generic dataset thinks is important.
- Per-practice-area tuning — via Corpus, you can teach the app what's "background" for your line of work. A med-mal corpus de-emphasizes medical boilerplate; a patent corpus de-emphasizes patent boilerplate. Same engine, different signal-to-noise.