Court Companion | AI Legal Multilingual Assistant

Pakistan’s AI-powered legal companion that makes the law & rights easy to understand through chat and voice support in 7 languages. Your rights. Your language. Your voice. Law made simple.

Comment

Court Companion | AI Legal Multilingual Assistant for Pakistan

Theme 3: Open Data & Access to Information

LIVE Here: https://ai-legal-assistant-seven.vercel.app/

| Admin dashboard | https://ai-legal-assistant-seven.vercel.app/admin |

| Admin Passkey | Admin123

Why I Built This

Here's the honest story behind this project.

At some point, I — along with some people I know — had an unsettling experience during travel. We were stopped and questioned in a way that felt intimidating. The officer brought up legal sections, offence numbers, and procedural language that I didn't fully understand. That moment was confusing and frightening. You just want to resolve the situation and move on — but you don't know what your rights are, what those law sections actually mean, or how to respond calmly.

I'm not making accusations against anyone — many officers serve with integrity. But that experience stayed with me. It showed how easily a lack of legal literacy can leave ordinary people feeling powerless, even when they've done nothing wrong.

When I saw the AI for Civic Innovation Hackathon, it clicked: what if legal rights and procedures were accessible to anyone, in their own language, on their phone — before or during a stressful moment?

That's why Court Companion exists.

Disclaimer: Court Companion provides legal information only — not legal advice or representation.

The Problem

Most citizens in Pakistan can't easily understand criminal law — FIR procedures, bail rights, PPC/CrPC sections — especially when they're under pressure, communicating in Urdu or a regional language. And real situations are rarely simple "What does Section X mean?" questions.

They look more like:

  • "The police are refusing to file my FIR — is that even allowed?"
  • "My cousin was arrested without a warrant — what happens now?"
  • "I acted in self-defence but I'm being accused of murder — can I even get bail?" Court Companion is designed to answer these real-world, scenario-based legal questions by combining statutory retrieval (PPC, CrPC, ATA) with plain-language explanations grounded in the law.

These are messy, multi-part, real-life situations — mixed with conflicting advice from neighbours, police, and "a guy who knows a guy." Keyword search doesn't work here. Generic AI just makes things up. People are left more confused than when they started.

What I Built

Court Companion is an AI legal assistant that thinks through real situations — not just definitions.

Ask it a textbook question like "What is PPC 379?" and it'll explain it clearly. But ask it a real, messy situation — with multiple facts, conflicting claims, and an actual decision to make — and it walks through it properly: identifies which laws apply, explains the relevant procedure, and lays out what the person's actual rights and next steps are.

  • In plain, simple language
  • Grounded in real Pakistani statutes (PPC, CrPC, ATA) — every answer is backed by the actual law text, not guesses
  • Using RAG (Retrieval-Augmented Generation) — finds the relevant sections first, then reasons through the scenario using them
  • In 7 languages for text — Urdu script, Roman Urdu, Pashto, Punjabi, Sindhi, Balochi, and English
  • Voice in English and Urdu — speak your question and hear answers aloud (more voice languages as resources allow)
  • Three ways in from home: Ask in chat · Ask by voice · Court Companion Pro (beta info for lawyers)
  • Runs on Android (APK) and web (Chrome)

Court Companion is an AI-powered multilingual legal assistant that helps Pakistani citizens understand their rights, FIR procedures, bail, and criminal laws (PPC, CrPC, ATA) in plain language. Built with Flutter for Android and Web, users can ask questions by text or voice in seven languages. Behind the app, a FastAPI backend on Render uses Retrieval-Augmented Generation (RAG): non-English questions are translated to English for search across 983 indexed statute chunks, then the model explains the retrieved legal sources in the user's chosen language with live streaming and source citations.

It makes trustworthy legal information accessible when people need it most — not as a dictionary, but as someone walking through the situation with you.

What It Can Handle — Real Scenarios, Not Just Definitions

This is the part that's hard to capture in a feature list, so here are real examples of the kind of situations Court Companion is built to handle. These aren't simplified "textbook" questions — they're the kind of messy, multi-part, real-life situations people actually bring.

Scenario — "My cousin was arrested without a warrant"

My cousin was arrested yesterday evening from his home without showing any warrant. The police said it is related to a fight that happened three days ago where someone was injured but not killed. He has not been produced before any magistrate yet and it has been more than 24 hours. If the injury was serious but not fatal, can police still hold him without bail, and under which sections of CrPC and PPC should his family understand his rights regarding bail and production before a magistrate?

Why this matters

A simple FAQ bot can answer "what is PPC 302?" Court Companion is built to handle the version of the question that actually shows up in someone's life — with conflicting information, emotional stakes, multiple legal issues tangled together, and a real decision on the other end. That's the gap between "looks like AI" and "is actually useful."

Target Audience

Primary Users Secondary Users
General citizens — people seeking quick and simple legal guidance Legal aid organizations — groups providing free or low-cost legal help
Students — learners exploring basic legal awareness and rights NGOs — organizations working on legal awareness and social justice
Rural communities — users with limited access to legal resources Community support groups — local bodies assisting people with legal issues
Low-literacy users — individuals needing easy, simplified legal explanations Legal researchers — professionals studying laws, cases, and legal systems
Individuals under stress or intimidation — people needing urgent legal clarity in difficult situations Lawyers & advocates — future Court Companion Pro workspace (beta info in app today)
Women and first-time complainants — especially in harassment, domestic disputes, or property cases where not knowing the process means giving up before even starting —

How It Works — The Tech Stack

Before I commit to a final stack, I'm testing all these things to answer two honest questions — is this actually feasible, or not? Am I on the right track, or not? Everything here is still being evaluated. Later on, the plan is to replace proprietary APIs with open-source alternatives where possible.

Step-by-step (what happens when you ask a question)

Step Layer What happens
1 User Opens Ask in chat or Ask by voice from home; picks response language.
2 Flutter Sends POST /ask/stream with { question, language, device_id, conversation_id } to Render.
3 API Validates request; checks FAISS index and Together.ai API key.
4 API Resolves or creates a conversation (Supabase) for multi-turn context when configured.
5 API IF greeting/small talk → skip search, LLM replies briefly. ELSE → legal RAG path.
6 API IF question is not English → Llama 8B translates it to an English search query.
7 API Embeds the English query → FAISS retrieves top PPC/CrPC/ATA chunks (hybrid keyword + vector).
8 API Builds prompt: prior turns + English statute text + mandatory answer-language rule from app picker.
9 API Primary LLM streams the answer token-by-token in the user's chosen language (model is env-configurable — Gemma 4 31B on Render today).
10 Flutter Shows live text, top 3 source chips (most relevant), disclaimer; optional 👍/👎; voice mode speaks each sentence as it arrives.

Court Companion uses two separate language pipelines that must not be confused:

Pipeline Purpose When it runs
A — Search translation Turn the user's question into English so FAISS can find the right statute chunks Before retrieval
B — Response language Tell the LLM which language to write the answer in During generation

The knowledge base (embeddings + FAISS index) is English-only.
The user's answer can still be in any of 7 languages.

Knowledge Sources

Source Coverage
Pakistan Penal Code (PPC) Offences, punishments, sections
Criminal Procedure Code (CrPC) FIR, arrest, bail, investigation
Anti-Terrorism Act (ATA) Scheduled offences & procedures

Why RAG and not just a chatbot?

Legal statutes are long, dense, and multilingual — keyword search alone fails. Citizens describe real situations, not search terms. RAG retrieves the relevant sections first, then the LLM reasons through the scenario using them. The translate-before-retrieve step makes an English-indexed corpus work across Urdu, Pashto, Sindhi, and more — so the reasoning step still happens in the user's own language.

Technology Stack

Layer Technology
Frontend Flutter (Dart) — Android, Web
Backend FastAPI (Python)
Database Supabase PostgreSQL — conversations, messages, feedback, usage events
Vector search FAISS (file-based, 983 vectors)
LLM API Together.ai
Hosting Render (API) · Vercel (landing + admin UI)
Voice speech_to_text, browser TTS / flutter_tts
Admin React + Vite at /admin on landing site

Models — what we use and how

Model Provider Role in pipeline
meta/meta-llama/Meta-Llama-3-70B-Instruct-Turbo Together.ai Primary answer generation on deployed API (Render LLM_MODEL)
meta-llama/Meta-Llama-3-8B-Instruct-Lite Together.ai Query translation — non-English questions → English before FAISS search
all-MiniLM-L6-v2 sentence-transformers (local) Embeddings — runs on the API server; no per-query API cost

How the LLM is used (not a raw chatbot):

  1. Retrieve first — FAISS finds relevant PPC/CrPC/ATA chunks (hybrid vector + keyword).
  2. Constrain the prompt — system prompt says: answer only from retrieved context; cite sources; legal information not advice.
  3. Guard output — sanitize step strips unsafe formatting; preserve Urdu script.

The model never sees the full statute books — only the top-k retrieved chunks plus short conversation history (capped). That keeps token use bounded and reduces hallucination risk.

Model choice is swappable — change LLM_MODEL in Render env; no app rebuild required.

Cost — honest numbers

Layer Hackathon / pilot cost Notes
Render API $0 (free tier) Cold start ~30–50s after idle; fine for demo, not for 24/7 production
Vercel (landing + web) $0 (hobby) Static landing + Flutter web build
Supabase $0 (free tier) Conversations, feedback, usage events
Embeddings $0 at runtime Pre-built index; MiniLM runs locally on server
Together.ai Pay-per-token — main variable cost Only paid component at scale

Cost controls already in the architecture:

  • RAG reduces tokens — model reads ~5–8 chunks, not whole corpora
  • max_tokens caps per request (LLM_MAX_TOKENS on Render)
  • Conversation history limited (CONVERSATION_HISTORY_LIMIT=10)
  • Smaller 8B model for translation only; larger model only for final answer
  • Admin dashboard tracks volume — can throttle or switch to cheaper model if usage spikes

Scalability

Today (prototype): One Render instance, file-based FAISS in memory, Supabase for state. Good for hundreds to low thousands of users/month on free tiers.

What scales cleanly:

Piece Now Scale path
Client Flutter web + APK Stateless; CDN/Vercel handles web traffic
API Single FastAPI container Horizontal replicas behind load balancer
Vector search FAISS in RAM (~983 vectors) Chroma / pgvector / Pinecone when corpus grows to case law
LLM Together.ai API Managed inference — no GPU ops on our side
DB Supabase Postgres Connection pooling; read replicas at NGO scale
Index builds Offline build_index.py Rebuild on corpus update; no downtime if blue-green deploy

Features

Everything below is implemented and working in the current prototype — not a mockup.

App & access

Feature What you get
Home screen Ask in chat · Ask by voice · Court Companion Pro (beta)
Platforms Flutter web (Vercel — primary) + Android APK (release build → Render API)
Landing page React marketing site with Pro section, live links, admin at /admin
API status Online / offline badge on chat and voice screens
No login required Anonymous device_id — citizens use the app without accounts

Text chat

Feature What you get
Streaming answers Live token stream, markdown, source chips, disclaimer
7 text languages English, Urdu (script), Roman Urdu, Pashto, Punjabi, Sindhi, Balochi
Chat history sidebar Reopen past conversations; tap New chat for a fresh thread
Session follow-up Same chat thread keeps context — e.g. FIR first, then "What about bail?"
Document attach PDF/TXT via clip icon in chat or /analyze-document API
Answer feedback 👍 / 👎 on each reply → admin Helpfulness metrics

Voice assistant

Feature What you get
English voice Mic input (STT) + spoken streaming answers (TTS)
Urdu voice Speak in Urdu (اردو) and hear replies in Urdu script — same RAG backend as chat

Legal intelligence (RAG)

Feature What you get
Statute corpus PPC, CrPC, ATA — 983 indexed chunks with section metadata
Hybrid retrieval FAISS vector search + keyword signals; translate-before-retrieve for non-English
Scenario reasoning Multi-fact questions → applicable law, procedure, rights, next steps
Source citations Every legal answer tied to retrieved statute excerpts; UI shows top 3 most relevant chips
Output guard Sanitized LLM output; preserves Urdu punctuation and formatting

How analytics categorizes "hot topics":

  • When a user asks a question, the API runs keyword-based topic detection on the question text (e.g. fir, bail, theft, ppc_sections, arrest_rights, fraud, assault, terrorism, or other).
  • These tags are stored on usage_events → admin Top topics chart.
  • Feedback (👍/👎) is stored separately in answer_feedback — drives Helpfulness % (up vs down in last 7 days). Topics on feedback rows can be enriched later by linking to the parent question.

Dependencies & Feasibility

I want to be upfront about what I'm relying on — and what breaks if anything changes.

Dependency What I rely on Risk if it changes
Together.ai LLM + API for generation and translation Model deprecation → swap model name in .env; that's it
Render Free-tier HTTPS hosting for FastAPI + FAISS Cold starts; may need paid tier for uptime
Supabase Conversation history, feedback, analytics Service outage → local chat still works; server context may lag
sentence-transformers Local embeddings (all-MiniLM-L6-v2) Pinned version; index rebuilds if changed
FAISS index File-based vector store (983 chunks) Rebuilt from build_index.py if corpus changes
Flutter Cross-platform client (APK + web) Standard; no vendor lock-in

Is this easy to replicate?
A basic ChatGPT wrapper — yes, in days. My stack — curated legal RAG, 7-language pipeline, voice, conversation persistence, admin analytics, APK, deployed API — would take meaningful effort. The legal data curation and section-aware indexing are the hardest part to copy quickly.

What am I currently validating?
I'm not claiming a final model choice yet. I'm actively testing:

  • Answer quality across languages (sindhi, balochi)
  • Reasoning quality on multi-part, scenario-based questions
  • Multi-turn follow-up quality with conversation context

Known constraints (being honest):

  • No dedicated funding yet — relying on free / minimal-cost resources
  • First request after idle can take ~50 seconds while Render wakes up
  • Voice: English + Urdu live today; five more languages once STT/TTS resources are finalized
  • Agentic clarifying questions (AI asks you first) ship with Court Companion Pro, not free chat yet

Language Support

Mode Supported now Coming next
Text (chat) English, Urdu (script), Roman Urdu, Pashto, Punjabi, Sindhi, Balochi —
Voice (STT + TTS) English and Urdu (script) Roman Urdu, Pashto, Punjabi, Sindhi, Balochi — as device/API resources are finalized

What's Coming — Roadmap

Court Companion Pro (lawyers) - Paid subscription

  • Full case workspace — upload pleadings, FIR, orders; one case = full context throughout
  • Agentic follow-ups — e.g. "Someone took my phone" → AI asks Was it taken by force? Do you know the person? → then cites PPC + precedents
  • Case-law RAG from public Supreme Court, High Court, and Sessions judgments
  • Gap / procedural flags for counsel (limitation, bail category, evidence gaps)

Platform & scale

  • Voice for remaining languages — Roman Urdu, Pashto, Punjabi, Sindhi, Balochi (STT + TTS)
  • Horizontal API scaling — multiple Render/Fly instances + load balancer
  • Vector DB migration — pgvector or managed search when case-law corpus grows
  • Offline FAQ cache — core rights without internet (rural users)
  • Open source release — public repo, docs, contribution guide
  • Corpus expansion — constitutional, civil, family law

Quick Summary

Question Answer
Working prototype? Yes — web app, APK, chat, voice (EN + Urdu), 7 text languages, admin, deployed API
Live demo URL? Web: https://ai-legal-assistant-two.vercel.app/ · Landing: https://ai-legal-assistant-seven.vercel.app/
AI used meaningfully? Yes — RAG over 983 legal chunks + multilingual LLM pipeline + multi-turn context
Models & cost? Together.ai (Llama 70B generates + Llama 8B translate); ~$0 infra on free tier; pennies per question at pilot scale
Civic impact? Legal rights in plain language, 7 text languages + Urdu voice, on mobile; data-driven admin KPIs
What's next? Court Companion Pro workspace; more voice languages; case-law corpus; open-source path

Built by: Syed Shah Hussain Badshah — June 2026 | AI for Civic Innovation Hackathon 2026 — Code for Pakistan × Grey Software × Scrimba × FAST NUCES — Final Round

Built With

  • all-minilm-l6-v2
  • apk
  • cursor
  • dart
  • faiss-index
  • flutter
  • github
  • llm
  • rag
  • render
  • stt
  • tts
  • voice
Share this project:

Updates

posted an update —

Chat streams answers in real time, with source citations and optional PDF/TXT document upload. Voice mode lets you ask in English (for now only adding more languages soon) by mic, with the answer spoken as it streams. The backend is deployed on Render; the Flutter app works on web and mobile.

Polishing voice for more languages and getting ready for wider release. Feedback welcome : )

Log in or sign up for Devpost to join the conversation.