Suit: Graphs, Vision, and the End of the "Friday Afternoon Filing"

Change Management, Not Mandate

The process for all filing should be only 3 parts: advocate enrollment number, Case ID (FR/CIN/CNR), and raw file dump. Done.

We don't need "mandatory e-filing" in the age of vision AI. OCR is deprecated. LLMs now handle physical scans, bookmarking, and docket prep with high fidelity. A physical AI-enabled filing desk with a simple scanner is perfectly good.

AI should be a Day 0 strategy for filing. The true value of a digital record isn't portability—it is memory and scrutability. A record the machine can read is a record the institution can remember across vacations, roster shuffles, and decades.

The Pipeline Gap: Where Time Actually Goes

The e-filing debate fixates on eliminating the trip to the registry. But that's 5 minutes of a 2-hour workflow. What happens after a docket is accepted? It goes through scrutiny: 150+ checklist items, HC rules, procedure, court fees deficiency against receipt against memorandum. Then defects are flagged and cleared. Only then does it get a case number.

After numbering comes docket prep: cover sheet, table of contents, consolidation of files, page numbering, bookmarking, and CIS metadata extraction—200+ fields. Then to the listing registrar for classification. Then to admission, where judges still lose case notes and navigate clunky dashboards mixed with paper submissions at the bench.

The registry spends 1–3 hours per case before listing. This kills time-sensitive matters—a bail application waiting on scrutiny is a day of liberty lost; a tender dispute filed at 3 PM may miss tomorrow's contract award. Machine scrutiny has no fatigue curve and catches more defects with complete audit trails.

The Technical Gap: Why RAG Fails at the Registry

Most legal AI products are built on RAG—Retrieval-Augmented Generation. RAG is how ChatGPT and its clones handle documents: chop the text into chunks, convert each chunk into a vector, store them in a database. When you ask a question, the system finds the "most similar" chunks and feeds them to the language model.

This works for research. Ask "find me cases about murder" and RAG retrieves relevant passages. The probabilistic nature is fine—you want breadth, not guarantees. But filing scrutiny is the opposite problem. The registry doesn't need the most likely location of a signature. It needs certainty that every signature, every annexure, every affidavit has been verified. RAG cannot provide completeness guarantees because it is architecturally incapable of promising it has examined everything.

The alternative is graph traversal. Instead of chopping documents into searchable chunks, we build a graph: the filing is a tree of documents, each document is a sequence of pages, each page contains elements (text, signatures, seals, photos). Sub-agents traverse every node. The system can prove it has touched every page, every margin, every annexure—because the graph structure makes omission detectable.

What Suit Does

Suit builds a graph representation of both the legal logic and the physical document structure. At inference time, sub-agents traverse every node—every page, every margin, every annexure—performing formal verification. The graph makes omission detectable: we can prove the AI has examined everything.

We use text extraction, OCR, vision, graph traversal, and legal search together. Text and OCR handle printed content. Vision AI handles what OCR misses: photos, diagrams, handwriting, signatures, seals, scan artifacts. The graph connects documents to each other—main petition to annexures, lower court record to appeal memo. Legal search finds precedent and statutory context. These are additive capabilities.

This allows us to handle arbitrary context length. We coordinate as many sub-agents as needed to map the graph, whether the file is 10 pages or 10,000.

Process Change

The computer can now assist with everything from filing scrutiny to TOC prep and PDF merging. It retains notes across vacations and roster shuffles.

The practical impact: lawyers get instant defect feedback before filing. Registry staff save 30–90 minutes per case on mechanical verification. The same metadata powers listing classification and case dashboards downstream.

The design is Human-in-the-Loop. Every AI output routes to a dashboard for staff review. The machine proposes; the human disposes. This is how jhana Courtroom works—registry staff retain judgment calls while the machine handles mechanical checks.

Classification & Related Matters

Filing is only the first step. The real downstream value comes from what the system learns about each case—and how it connects cases to each other.

One consequence: forum shopping becomes visible at filing. When the same relief is sought in multiple forums, or related matters are pending before different benches, the graph surfaces this. Judges can tag and link related proceedings; conflicting orders are caught before they issue. The docket becomes legible to the institution.

Time-Sensitive Matters: Where Speed is Justice

Some matters cannot wait for next-day scrutiny. The current filing infrastructure fails precisely where urgency is highest.

For tender matters, bail disputes, and any hearing where a day's delay changes outcomes, the filing bottleneck is the justice bottleneck.

What Users Are Saying

Users deploy Suit for summary, translation, diligence, and discovery. A few examples:

I used Suit to redline a contract and it immediately mapped the LD clause to the relevant ONGC jurisprudence.

— Pranav Menon, LinkedIn

Suit's annotations of issues and prayers in a docket are sharper than the counsel's.

— A sitting Judge

The "Friday Afternoon" Filing

The filing story is about litigants. It always has been.

Your opposite party has always played dirty:

• Friday afternoon filings of nonsense papers to get a Monday date.
• Globally standard 3-page contracts that become essays and side letters.
• 200-page dockets filed to court relying on the idea that the judge cannot read them.

We rely on the inscrutability of the legal record as a process advantage. Fattening the file is a strategy. What happens when the record becomes instantly scrutable? The fraudulent Friday filing gets rejected with a defect list on Monday morning. The 200-page docket gets read.

Machine Scrutiny of Filing

Can we do machine scrutiny of filing—both quick e-filing and physical filing which is scanned? Yes. As long as we have the digitized file (born-digital or scanned), the graph doesn't care about the source.

All in less than 15 minutes, with audit trails.

The machine catches more true-positive defects than physical scrutiny. Consistency at scale is what computers do. Human expertise is preserved for judgment calls; mechanical checks are automated.

This sidesteps the mandatory e-filing debate. AI-enabled quick filing kiosks with simple scanners produce the same output as born-digital submissions: a verified, scrutable, complete record.