Multi-Agent Legal Research Pipeline
Abrar Mohtasim
ABRAR MUHTASIM
Agentic AI Engineer • LLM Systems Builder • Multi-Agent Architect
I build AI systems that reason, route tasks, call tools, and deliver grounded outputs
Agentic AI • Multi-Agent Orchestration • Tool Calling • LLM Fine-Tuning • AI Workflow Automation
About Me
I am an Agentic AI Engineer who builds production-grade AI systems where LLMs go beyond answering questions — they plan, route tasks, call external tools, manage conversational context, verify outputs, and automate real business workflows.
My strongest work is in:
Agentic AI Engineering — designing multi-agent pipelines with role specialization, sequential orchestration, tool use, and query routing
LLM Engineering — prompt design, output grounding, anti-hallucination systems, fine-tuning, and evaluation
AI Product Engineering — turning agent prototypes into deployable systems with databases, dashboards, APIs, and automation layers
I care about building AI that is not just impressive in demos — but reliable, verifiable, and genuinely useful in production.
Core Engineering Focus
🧠 Agentic AI Engineering
Multi-agent orchestration with specialized agent roles
Sequential context chaining across agent handoffs
Query routing and intent-based flow control
Tool-calling agents using ReAct and native function calling
Conversational memory using session and cache layers
Anti-hallucination verification at the system level
⚙️ LLM Systems Engineering
LangChain, CrewAI, LiteLLM, OpenRouter
Structured outputs, function calling, and tool schemas
Redis for session memory and response caching
Citation-grounded and retrieval-augmented generation
QLoRA fine-tuning, dataset design, and model evaluation
Gradio, Streamlit, HuggingFace Spaces, Render deployment
Featured Projects
Automated Case Law Retrieval, Statute Analysis and IRAC Memo Generation
Takes raw client facts as input and outputs a fully structured, citation-verified legal memorandum — built for legal professionals who need reliable, grounded research at speed.
Designed a 5-agent CrewAI pipeline where each agent has a distinct role: intake, legal retrieval, statute analysis, synthesis, and citation verification — with structured output passed sequentially between agents
Implemented query routing to classify incoming requests into full research, follow-up, refinement, or clarification flows before any agent is invoked
Built a zero-trust citation system where no legal reference enters the final memo unless it has been retrieved and confirmed through CourtListener or Tavily tool calls
Engineered a 4-layer output verification system to catch hallucinated citations, unsupported claims, and incomplete IRAC structure before final delivery
Skills Demonstrated: Multi-agent orchestration, tool calling, query routing, output grounding, memory management, anti-hallucination design
CrewAI LangChain Redis OpenRouter CourtListener API Tavily Gradio Domain Fine-Tuned LLM on SEC Filings with Hallucination Resistance
Takes a financial question and an SEC filing excerpt as input and outputs a cited, evidence-grounded answer — or an explicit refusal when the filing does not contain sufficient evidence.
Fine-tuned Phi-3-Mini (3.8B) using QLoRA + 4-bit quantization — completed in approximately 25 minutes at zero compute cost on free-tier GPU
Designed the training dataset with a 70/30 split — 70% matched question-context pairs and 30% deliberately mismatched pairs — to teach the model when NOT to answer
Improved citation inclusion rate from 80% to 100% by making citation a structural requirement enforced through prompt format and training signal
Improved answer rate from 45% to 100% on matched contexts while preserving correct refusal behavior on mismatched ones
Trained and deployed a live HuggingFace Spaces demo with Gradio interface for interactive testing
Skills Demonstrated: QLoRA fine-tuning, dataset engineering, hallucination resistance through training, model evaluation, HuggingFace deployment
Phi-3-Mini QLoRA BitsAndBytes TRL SFTTrainer HuggingFace Gradio Natural Language to SQL Interface Over 24 Hotel Performance KPIs
Takes plain-English business questions as input and outputs SQL-backed KPI answers, anomaly alerts, and property health scores — built for hotel operations teams who need analytics without writing queries.
Built an LLM agent with native function calling that routes natural language questions to specific SQL tool functions rather than generating raw SQL directly — preserving control and reliability
Engineered a deterministic SQL engine covering 24 business KPIs including RevPAR, occupancy rate, ADR, cancellation rate, and booking channel mix — all with pre-validated query logic
Implemented a 4-layer anomaly detection system that flags statistical outliers in KPI data and assigns property-level health scores for executive reporting
Unified the agent answer layer and the Streamlit dashboard through a single shared metrics engine — ensuring the agent and the dashboard always show the same numbers
Used Supabase + PostgreSQL as the backend with Plotly for interactive chart rendering inside the Streamlit interface
Skills Demonstrated: Function calling, tool schema design, deterministic SQL pipelines, anomaly detection, dashboard engineering, agent-UI integration
Python LiteLLM OpenRouter PostgreSQL Supabase Streamlit Plotly Impact at a Glance
100%
Citation Rate
Improved from 80%
100%
Answer Rate
Improved from 45%
~90%
Cost Reduction
Via model routing
4
Verification Layers
Anti-hallucination
$0
Fine-Tuning Cost
Free-tier GPU
25 min
Training Time
Phi-3-Mini QLoRA
24
Deterministic KPIs
SQL-backed metrics
2
Publications
IEEE + Acadlore
Technology Stack
Agentic AI and LLM Frameworks
Fine-Tuning and Machine Learning
Memory, Data and Infrastructure
Apps, Deployment and Automation
Additional Projects
🎁 BloomGift AI
AI-powered gift and event recommendation platform that personalizes suggestions based on recipient profile and occasion type.
Lovable Supabase OpenRouter💪 ThrivePath
AI wellness assistant that generates personalized meal plans and fitness routines based on user goals and dietary preferences.
Lovable Supabase OpenRouterWorkflow Automation
Built production AI automation workflows using n8n including automated email triage and an AI voice agent for appointment scheduling using ElevenLabs and Google Calendar API
Analytics and Data Projects
E-Commerce Web Analytics — SQL analysis surfacing 11% CTR improvement opportunity and 40% cart abandonment pattern
Customer Segmentation and Churn Analysis — K-means and RFM segmentation identifying 23% churn-risk customer cohort
Hotel Revenue Insights — Power BI dashboard revealing 32% cancellation rate and 52% weekend occupancy patterns
Hotel Booking Cancellation Prediction — Random Forest classifier achieving 78.34% accuracy on booking cancellation prediction
Research and Publications
Optimizing Predictive Models for Drug-Induced Suicidal Risk — IEEE Xplore, 2025
GitHub Analytics
Education and Certifications
B.Sc. in Computer and Communication Engineering International Islamic University Chittagong — CGPA 3.39 / 4.00
Certifications
365 Data Science — Data Analyst Career Track
Data Science using Python A-Z for Machine Learning
PostgreSQL Bootcamp: Beginner to Advanced
IELTS — Overall Band 6.5
Currently Exploring
Advanced agentic workflow patterns including parallel agent execution and hierarchical delegation
Production memory design for long-horizon multi-turn AI conversations
Fine-tuning and alignment strategies for domain-specific reliable outputs
AI workflow automation at scale with n8n and API integration layers
Let's Connect
Open to Agentic AI Engineer, LLM Engineer, AI Product Engineer, and Applied ML roles
Like this project
Posted Apr 26, 2026
Developed a multi-agent legal AI pipeline for reliable legal research and citation verification.
Likes
0
Views
0
