Explore

Projects - Aditya Goel

1. Adira AI

Purpose: Helps users draft any legal document - contracts, notices, affidavits, and more instantly, with zero legal formatting errors.

Key Capabilities:

Converts user input into full-length legally valid documents

Edits, redlines, and auto-corrects clauses

Suggests missing legal sections based on document type

Behind the Scenes:

Operates on a legal language model fine-tuned on thousands of Indian contracts

Maintains template logic, jurisdictional nuance, and conditional phrasing

Smart handling of variable fields like party names, dates, etc.

Tech Stack / Tools Used:

Python (Core)

LLM APIs (OpenAI GPT-4 / fine-tuned legal LLMs)

python-docx, pdfplumber, reportlab for DOCX/PDF output

re (regex), SpaCy for clause detection and parsing

Internal template management system for dynamic field handling

Business Outcomes Achieved:

Reduced average drafting time from 2–3 hours to under 10 minutes

Enabled over 5000+ documents auto-generated in alpha phase

Achieved 98% accuracy in clause structure and jurisdictional tagging in pilot feedback

Enabled law firms to serve 2.5x more clients per day with the same team

Optimizations and Their Outcomes:

Replaced static prompts with dynamic templates → improved personalization and formatting accuracy by 40%

Batched clause suggestion engine → reduced processing time per document by 25%

Added “smart fill-in” placeholders for key terms → reduced manual editing post-generation

Link :

www.adiralaw.com⁠

⁠

2. AI Courtroom

Purpose: Simulates real courtroom hearings to prepare lawyers with case analysis, counterarguments, and judgment predictions.

Key Capabilities:

Auto-generates opposing counsel arguments

Highlights weak points in uploaded petitions or briefs

Suggests case laws and potential judge opinions

Behind the Scenes:

Multi-layer AI logic for each actor in the courtroom (Lawyer, Opposing Lawyer, Judge)

Contextual understanding of petitions and legal issues

Generates dynamic legal rebuttals and closing arguments

Tech Stack / Tools Used:

Python

LLM API orchestration using memory-based contextual chains

PyMuPDF, pdfminer for parsing legal PDFs

Text classification using scikit-learn for weak point detection

Precedent case fetch logic using indexed search of Indian judgments

Business Outcomes Achieved:

Reduced hearing prep time from 7–8 hours to ~15 minutes

Simulated over 700+ courtroom rehearsals in pre-release trials

Helped litigation teams improve argument completeness by 60%, as measured by missing citations and issue handling

Optimizations and Their Outcomes:

Cached key phrase interpretations → reduced repeat processing by 40%

Shifted to staged context loading for memory efficiency → enabled longer case briefs to be processed without lag

Layered judge logic with precedent lookup → improved judgment prediction relevance by 33%

Link :

www.courtroom.clawlaw.in⁠

⁠

3. LegalGPT

Purpose: Acts as an internal knowledge base across products - answering legal questions, summarizing judgments, and interpreting statutes.

Key Capabilities:

Conversational Q&A based on Indian law

Automatic case summarization

Highlights relevant sections and links laws to facts

Behind the Scenes:

Continuously trained on Indian case law, statutes, and regulatory updates

Integrated across other products to support context-aware decision-making

Tech Stack / Tools Used:

Python

Fine-tuned LLM using Indian legal datasets (case law, statutes, bar council updates)

Retrieval-augmented generation (RAG) over indexed legal corpora

transformers, sentence-transformers, FAISS for semantic similarity and response ranking

Scheduled refresh of law databases via internal ETL sync

Business Outcomes Achieved:

Used as a central engine across Adira and Courtroom tools

Reduced legal research effort by 70% per query for law firm clients

Delivered over 10,000+ responses with domain-aligned legal citations and summaries

Increased team-wide response consistency for legal support teams by 50%

Optimizations and Their Outcomes:

Fine-tuned answer length control → reduced verbosity and improved response quality

Legal topic classifier → improved answer routing speed by 20%

Embedding model upgrade → improved semantic hit rate by 35% for user queries

Link :

www.legalgpt.clawlaw.in⁠

⁠

4. Legal ETL Engine

Purpose: Automates the end-to-end process of collecting, cleaning, and updating Indian Acts into our legal database.

Key Capabilities:

Scrapes Indian laws from trusted government sources (e.g., India Code)

Cleans formatting, extracts sections and sub-sections

Regularly adds or updates central and state laws in the internal index

Behind the Scenes:

Built as an event-driven pipeline using AWS Lambda functions

Scheduled extraction, transformation, and load workflows

Data normalization ensures uniformity in sectioning, tagging, and clause referencing

Automatically flags amendments or new acts for human legal review before indexing

Tech Stack / Tools Used:

AWS Lambda, AWS S3

Python

requests, BeautifulSoup, lxml for scraping

pandas for data wrangling

re, nltk, spaCy for legal clause structuring and token-level cleanup

Business Outcomes Achieved:

Automatically ingests and updates over 200+ Indian Acts and amendments monthly

Eliminated 100% of manual legal data entry tasks

Improved clause mapping accuracy across products by 40% due to structured legal indexing

Enabled enterprise clients to access up-to-date statutes with 0 lag time

Optimizations and Their Outcomes:

Migrated to lightweight Lambda triggers → reduced compute cost by 60%

Implemented delta-based update checks → minimized redundant downloads and reprocessing

Added a semantic flagging layer → improved Act amendment detection accuracy by 45%

⁠

5. Lawyer Connect

Purpose: This system was built to power Claw’s "Lawyer Connect" initiative - a network feature designed to match lawyers with other lawyers (for referrals, collaborations) and clients (for legal services) based on availability, expertise, ratings, and past behavior.

Key Capabilities:

Connects two user types: Clients and Lawyers / Lawyer to Lawyer (Referrals)

Recommends best-fit matches based on case category, language, location, and lawyer history

Enables dynamic re-ranking based on real-time availability, communication score, and case complexity

Supports feedback loop through user interactions to improve future recommendations

Behind the Scenes:

The recommendation engine uses a hybrid model combining content-based filtering (for skill and location) and behavioral data (for engagement, completion rate, responsiveness).

A modular architecture separates Lawyer-to-Lawyer and Lawyer-to-Client matching logics but uses a common embedding layer for user profiles.

Profiles are vectorized using multi-feature embeddings (experience, practice area, ratings, average response time, etc.), and nearest neighbors are identified using cosine similarity.

Real-time triggers update the recommendation scores when a lawyer accepts/declines or interacts with a profile.

Custom weight tuning was done to prioritize newer lawyers during off-peak times, helping equalize exposure.

Tech Stack / Tools Used:

Python

scikit-learn, NumPy for model logic

FAISS for fast nearest neighbor search

pandas for user profiling and ranking tables

Simple in-memory cache layer for recent interactions

CSV-based mock data ingestion layer for offline testing

Optimizations and Their Outcomes:

Replaced cosine similarity with hybrid score (distance + feedback weighting) → improved match relevance in A/B testing

Integrated response-time decay function → improved quality of active lawyer matches

Added asynchronous job queue for profile score updates → reduced response latency by 50%

Want to print your doc?
This is not the way.

Try clicking the ··· in the right corner or using a keyboard shortcut (

CtrlP

) instead.