J

Judicra

AI legal research for Indian law.
Trained to reason like a senior advocate.

Pre-seed Private beta judicra.com
The Problem

Legal research in India is broken.

1.8 million advocates. Most still rely on manual keyword search across fragmented databases, or generic AI that hallucinates citations.

1

Fragmented, expensive databases

Lawyers juggle SCC Online, Manupatra, Indian Kanoon, bare act sites. Annual subscriptions run INR 15,000-50,000+ each, with no synthesis across sources.

2

Keyword search, not legal reasoning

Existing tools return document lists, not answers. Lawyers spend 2-4 hours per research task manually reading and synthesizing judgments.

3

Generic AI fails on Indian law

ChatGPT and Gemini hallucinate case names, fabricate section numbers, and miss India-specific codes (BNS, BNSS, BSA replaced IPC, CrPC, IEA in 2024).

Market Opportunity

India's legal services market is massive and underserved by technology.

$15B
Indian legal services market
$1.2B
Legal tech TAM (research + workflow)
$350M
SAM: AI-powered legal research

1.8M

Registered advocates in India (Bar Council)

50M+

Pending cases across Indian courts

~70%

Solo practitioners or 2-3 person firms

3 new codes

BNS, BNSS, BSA (2024) created urgent reskilling demand

Initial SOM: 50,000 solo/small-firm advocates at INR 500-2,000/mo = INR 30-120 Cr ARR ($3.5-14M)

The Solution

Ask in plain language. Get a senior advocate's analysis with citations.

What are the essential ingredients of a Section 138 cheque bouncing offence?
J
To establish an offence under Section 138 of the Negotiable Instruments Act, 1881, the complainant must show: Cheque drawn by the accused on an account maintained by them. Legally enforceable debt or liability for which the cheque was issued. Presentation within validity of the cheque. Dishonour by the bank for insufficiency of funds. Written demand notice within 30 days of receiving information of dishonour. Section 139 raises a rebuttable presumption that the holder received the cheque towards a debt or liability.

Not a chatbot wrapper. A purpose-built legal AI.

  • Structured answers with bold section numbers, case names, and provisions, ready for a brief or memo.
  • Verifiable citations grounded in 5M+ real court judgments, not hallucinated.
  • Conversational follow-ups. Refine your research in natural dialogue, not keyword reformulation.
Technology

Custom model + comprehensive legal corpus.

Fine-tuned LLM

Judicra 1.0: Qwen3.5-27B fine-tuned with LoRA on 20,000+ expert-curated Indian legal QA pairs. Trained to reason with Indian statutes, case law, and the new criminal codes.

Not a prompt wrapper over GPT/Gemini.

RAG over 10M+ vectors

5M judgments from SC, all 25 HCs, 16 tribunals, 4,000+ statutes. BGE-M3 dense+sparse embeddings with cross-encoder reranking. Every answer is grounded in real documents.

Hybrid retrieval + reranking = precision.

Anti-hallucination by design

Adversarial training data (fake provisions, trick questions). Citation honesty training. Knowledge-grounded refusal for questions outside the corpus. The model says "I don't know" when it should.

601 adversarial + 382 citation QA pairs.
5M+
Court judgments indexed
10.3M
Vectors in production
20K+
Training QA pairs
~19 tok/s
Streaming generation speed
Defensibility

Compounding moats, not features.

Proprietary training data

17 custom QA generation pipelines. Each pipeline has domain-specific validation, fact-checking, and cross-referencing. This data doesn't exist anywhere else. Every iteration improves the model.

India-specific corpus at scale

5M+ judgments across all Indian courts, cleaned, deduplicated, and quality-scored through a 5-phase pipeline. Coverage that took months to build and would take competitors the same.

Vertical model advantage

General-purpose LLMs fail on Indian legal nuance: new criminal codes, tribunal-specific procedures, state-specific amendments. Our fine-tuned model handles these natively.

Usage data flywheel

Every query improves retrieval quality and reveals gaps in training data. Server-side audit logging captures real practitioner questions for targeted model improvement. The product gets better with every user.

Landscape

A different approach to Indian legal research.

Judicra Incumbent databases Generic AI (ChatGPT, etc.)
Answer format Synthesized analysis with citations Document list (user synthesizes) Prose answer (often wrong)
Indian law depth Fine-tuned on Indian legal corpus Comprehensive but search-only Superficial, hallucinates specifics
New codes (BNS/BNSS/BSA) Trained with crosswalk mappings Available as raw text Frequently confuses with old codes
Citation reliability RAG-grounded, adversarial-tested Real citations (search results) Fabricates case names and sections
Pricing INR 500-2,000/mo INR 15,000-50,000+/yr per DB $20/mo (not specialized)
UX Conversational, instant 1990s search interface Good UX, wrong answers
Business Model

SaaS with high-margin AI delivery.

Free Trial

Free

Limited queries. Conversion funnel to Pro.

  • Limited daily queries
  • Full model quality
  • No conversation history

Pro

INR 999/mo

~$12/mo. Target: solo practitioners.

  • Unlimited queries
  • Full 27B model
  • Conversation history + export
  • Priority support

Firm

INR 2,499/seat/mo

~$30/seat. 3+ users.

  • Everything in Pro
  • Team workspace
  • Report generation
  • API access

Unit economics: GPU inference cost ~$0.002/query (serverless scale-to-zero). At INR 999/mo and ~100 queries/user/mo, gross margin exceeds 85%.

Expansion revenue: Firm tier, API access for litigation support tools, report generation credits, specialized tribunal modules.

Go-to-Market

Bottom-up adoption, starting with solo practitioners.

Phase 1: Solo lawyers + small firms

  • Product-led growth. Free trial drives organic adoption. Share button on every answer turns users into distributors.
  • Bar association partnerships. Subsidized or free access through state bar associations, CLE programs.
  • Content marketing. Weekly legal analysis posts targeting high-volume search queries (anticipatory bail BNSS, Section 138 NI Act, etc.).
  • WhatsApp/Telegram groups. Indian lawyers live in these. Direct, peer-to-peer distribution.

Phase 2: Expand vertically

  • CAs and Company Secretaries. 400,000+ professionals who need tax tribunal orders, Companies Act compliance, SEBI/RBI circulars.
  • Law students. 1,500+ law colleges, 200K+ students. Discounted plans for campus adoption.
  • In-house legal teams. Mid-market companies (fintech, real estate, manufacturing) with 1-5 person legal teams.

Phase 3: Platform

  • API for legal tools. Litigation management software, compliance platforms, and contract review tools integrate Judicra's research engine.
Traction

What's been built so far.

Corpus assembled

5M+ judgments scraped, cleaned, deduplicated. 10.3M vectors indexed in Qdrant.

Model trained

Judicra 1.0 (27B) fine-tuned. 17 QA pipelines, 20K+ training pairs. Eval score: 6.4/10 (with RAG).

Product live

judicra.com in private beta. Full-stack: streaming chat, auth, conversation sync, PDF export, feedback loop.

Infrastructure production-ready

RunPod serverless (scale-to-zero), Cloudflare Pages, Mac Mini backend. <$50/mo running cost.

What's been built

  • 5-phase data processing pipeline with quality scoring
  • 17 QA generation pipelines with adversarial validation
  • 212-question eval suite with automated scoring
  • RAG system: hybrid dense+sparse retrieval with reranking
  • Fine-tuned 27B model with LoRA, deployed on serverless GPU
  • Full-stack product with SSE streaming, auth, sync
  • Server-side audit logging for model improvement

Current eval: 6.4/10

Target: 9/10. Biggest gaps: citation quality (4.8), new criminal codes (4.7), adversarial robustness. Clear path to improvement with more training data + RAG tuning.

Roadmap

Paid product launching May 2026.

Now - May 2026

Paid launch. Free trial + Pro live.

  • Judicra 2.0 training (expanded data, citation focus)
  • New criminal codes coverage gap closed
  • Self-serve onboarding, payment (Razorpay)
  • Feedback loop: audit logs to training data

Q3 2026 (Jun - Sep)

Growth. First 500 paid users.

  • Bar association partnerships
  • Report generation feature
  • Model iteration from user feedback
  • Target: 500 paying users by end of Q3

Q4 2026 - Q1 2027

Scale. Firm tier. API.

  • Firm tier with team workspaces
  • API access for integrations
  • CA/CS vertical expansion
  • Target: 5,000 paying users, INR 50L+ MRR

Key milestone: Eval score from 6.4 to 8+/10 is the unlock. At 8/10, the product is reliable enough for paying practitioners. Training data pipeline and eval suite are already built. The bottleneck is compute for iterative training runs and corpus expansion.

Team

Founding team.

Aditya Patni

Founder

End-to-end builder across the stack: data pipelines, model training, RAG, backend, frontend, and infrastructure.

aditya@judicra.com

LinkedIn

First hires (with funding)

Legal Domain Expert (part-time)

Training data quality, eval curation, practitioner feedback loop. Someone who practices law and can judge answer quality at an expert level.

Growth / GTM

Bar association partnerships, content marketing, community building. Someone who knows how Indian lawyers discover and adopt tools.

The Ask

Raising pre-seed to get to product-market fit.

Pre-seed round

$500K

18-month runway to paid product with 5,000 users

Use of funds

GPU compute (training + inference)

60%

Team (part-time legal expert + GTM)

20%

Infrastructure + tools

10%

Go-to-market spend

10%

Milestone for next round: 5,000 paying users, INR 50L+ MRR, eval score 9/10, Firm tier live.

J

Indian legal research,
reimagined.

The infrastructure is built. The model works. Now it's time to scale.

Try it

judicra.com

Contact

aditya@judicra.com

1 / 14