About

Why this project exists.

I hold an LLM in criminal law and I'm currently completing an MSc in Business Analytics. My dissertation examined the unequal application of criminal law in India; a related project of mine analyses consistency in UK asylum appeal outcomes using open tribunal data. The Liberty Model is the next step in that line of work — applying the same combination of legal reasoning and data methods to one of the starkest access-to-justice problems I know of: undertrial detention in India.

This page sets out, deliberately and in full, what this project is and is not.

What the data actually is

Synthetic case data (used in the dashboard)

The 120 cases shown in the triage dashboard are entirely synthetic — generated programmatically to resemble realistic distributions of offence type, custody duration, age, and prior record. No real individual's case data is displayed anywhere on this site. This dataset exists to demonstrate that the rules engine works correctly, not to make any claim about real cases.

Real aggregate statistics (used for context only)

Figures such as the "~75% of prisoners are undertrials" statistic on the homepage come from published NCRB Prison Statistics and NJDG (National Judicial Data Grid) aggregate data. These describe the scale of the systemic problem; they are not linked to any individual case shown in this demo.

Real case extraction pipeline (built, not yet deployed here)

Separately, I built a pipeline to extract case metadata from public eCourts/NJDG case-history pages for a small number of cases, as a proof of concept for how this system could eventually connect to real, anonymised court data under proper legal and data-protection review. That pipeline is not connected to this public dashboard, and no data it collected is published here.

What this system does not do

It does not predict reoffending risk, flight risk, or any other forward-looking judgment about a person.
It does not make or automate any legal decision. Every flag is a prompt for human legal review.
It does not use machine learning to derive its rules — every rule is a direct, traceable encoding of a specific statute or precedent, listed on the Methodology page.
It has not been validated on real case data, deployed with any court, prison authority, or legal aid body, or reviewed by a practising advocate. It is, at this stage, a research and portfolio demonstration.
It is not a substitute for legal advice, and nothing on this site should be relied upon as such.

What would need to happen before real-world use

Before any version of this tool could responsibly support real legal aid work, it would need: verification of every rule by a practising criminal lawyer against current law; a real, anonymised dataset with proper consent and data-protection safeguards; a pilot run alongside — never in place of — an existing UTRC or DLSA process; and an independent fairness audit of the flagging system on real case outcomes, not synthetic data. None of that has happened yet. This is a first step, not a finished product.

Feedback from legal aid practitioners, academics, or anyone working in this space is genuinely welcome — details on my LinkedIn profile.