Featured work

Legal Chatbot

AI-powered legal assistance and a virtual courtroom for Indian practice, built as an agentic RAG system: retrieval and grounding before the model ever answers.

Llama 3.3 70B (Groq)Ollama, local Llama 3 / Phi-3Deepgram STTAgentic RAGLangGraphCrewAIFastAPISQLitePyAnnote Group project · 3

Project repo Architecture View slides

Architecture

Four entry points — a typed question, voice, an uploaded document, or the virtual courtroom itself — all funnel into one retrieval-and-reasoning pipeline. Legal text is retrieved and ranked before Llama 3.3 70B ever sees it, so answers are grounded in the actual corpus rather than the model's memory.

1 Example walkthrough

User Interaction

Multiple ways to interact

Text Query

Voice Input

Document Upload

Virtual Courtroom

Input Processing

Convert and prepare input

Deepgram STT

Query Processing

Document Processing

Chunking

Legal Knowledge & Retrieval

Find the most relevant legal text

Legal Corpus

Vector / Retrieval Index

Retriever

Retrieval Ranking

AI Reasoning

Generate a grounded response

Context Construction

Llama 3.3 70B (Groq)

Output

Deliver the result

Legal Explanation

Document / Clause Analysis

Virtual Courtroom Response

Response delivered to user

As text, voice, or inside the virtual courtroom

User interaction Input processing Legal retrieval AI reasoning Output

How it works

01 — User interaction: a text query, voice input, an uploaded document, or the virtual courtroom interface — four entry points into the same backend.
02 — Input processing: voice goes through Deepgram speech-to-text first; documents are processed and chunked; everything else becomes a normalised query.
03 — Legal retrieval: the query is matched against the indexed legal corpus, and a retriever ranks the chunks that actually matter.
04 — Context construction: the query and the ranked legal chunks are assembled into one grounded prompt — the model never sees the raw corpus directly.
05 — LLM reasoning: Llama 3.3 70B, served through Groq, generates the answer from that grounded context.
06 — Response: delivered back as a legal explanation, a document/clause analysis, or a reply inside the virtual courtroom interface.
Three pipelines share this one engine: a research pipeline (about six agents that retrieve, draft and check), a courtroom pipeline (speaker diarisation with PyAnnote and live contradiction alerts), and a simplifier that turns legalese into plain language.

Model evolution

Development iteration — one model replacing the last as limits showed up, not several models running together in production.

01 Phi-3 (local) Could not reliably hold the multi-step agentic flow across a full session.

02 Smaller local Llama Held the flow, but took minutes per answer — too slow for a live legal session.

03 Llama 3.3 70B via Groq Moved the heavy reasoning to Groq's hosted 70B model, bringing answers down to 90–180 seconds.

My contribution

Key features

RAG-based legal query answering
Document upload and clause analysis
Voice interaction via Deepgram STT
Virtual courtroom interface
Grounded, citation-aware responses

Team

Legal Chatbot was built by a 3-person team as a group project, not solo.
Retrieval, agent orchestration, speech integration and the virtual courtroom interface were built together across the team.

What I worked on

RAG architecture and the retrieval workflow — the query-to-context path described in the architecture above.
Legal query processing and retrieval ranking.
Grounding and context construction for the LLM.
Model experimentation and the move from local models to Llama 3.3 70B on Groq.

Results

Phi-3 could not hold the agentic flow. A smaller local Llama worked but took minutes per answer. Moving heavy reasoning to Groq's 70B brought it down to 90 to 180 seconds, fast enough for a real session.
The team measured roughly a 70% cut in legal-research time on the research pipeline.

In code

python

ctx = retrieve(query, k=8)                    # local store, 500-char chunks
draft = llm(prompt(query, ctx))               # Groq Llama 3.3 70B, heavy reasoning
for claim in split_claims(draft):
    tag = "[FROM RAG]" if grounded_in(claim, ctx) else "[FROM WEB]"
    claim.annotate(tag)
answer = verifier_agent(draft, acts=ctx)      # conflict -> statute text wins▊

Grounding pass: every claim is tagged by source, then a verifier agent resolves conflicts.