Murtaza
Nipplewala

The Problem

Security Operations Center (SOC) teams process thousands of SIEM alerts every day. About 95% are noise, while the remaining 5% represent critical threats that cost millions per breach on average. SOC-Claw automates alert triage so analysts can focus on the threats that matter.

Approach — Why a second agent checks the first

A three-agent AI pipeline (Triage → Verifier → Response) that enriches raw alerts, double-checks its own reasoning, and proposes response plans that an analyst approves before anything runs.

Key Highlights

• Triage Agent enriches raw alerts with IP reputation databases, MITRE ATT&CK technique mapping, and asset CMDB lookups, producing P1–P4 severity scores with confidence ratings and reasoning chains
• Self-correcting Verifier Agent (no tools) runs a 4-point checklist — evidence alignment, reasoning completeness, logical consistency, and bias detection — improving triage accuracy from 78% to 88%
• Response Agent generates prioritized incident response plans with per-step reasoning and urgency levels, requiring analyst approval before execution
• Privacy-aware routing keeps sensitive SOC data (internal IPs, hostnames, alert payloads) on local inference via vLLM, with optional routing to cloud endpoints
• FastAPI backend and Red Hat-themed dashboard for real-time alert analysis, triage visualization, and per-step action approval or rejection

The Problem

Consumers often make impulsive purchases without understanding the true financial impact. Existing budgeting apps track spending but don't proactively evaluate whether a purchase is wise. SavVio bridges this gap by combining AI reasoning with deterministic financial rules to deliver personalized Buy/Wait/Avoid recommendations.

Approach — Why I didn't just use an LLM for everything

An AI-driven financial advocacy tool that evaluates purchase decisions using a hybrid architecture combining LLM-based context understanding with a deterministic financial logic engine for affordability analysis.

Key Highlights

• RAG pipeline with LangChain and pgvector product embeddings for context-aware product utility analysis through GPT-4.1, Claude 4.5, and Gemini 3
• Trained, tuned, and compared classifier and regression models for Buy/Wait/Avoid recommendations, with MLflow experiment tracking and model versioning
• Apache Airflow pipelines for data ingestion and model training, with Great Expectations data validation and Evidently AI drift monitoring
• NVIDIA NeMo Guardrails as FastAPI middleware for pre- and post-LLM response validation
• Containerized FastAPI services on GCP Cloud Run, with CI/CD through GitHub Actions and Cloud Build
• Prometheus, Grafana, and GCP Cloud Monitoring for latency, drift, and cost tracking with billing alerts, plus DVC data versioning on GCP Cloud Storage

The Problem

Asking one LLM prompt to plan a topic, explain it, and quiz you on it mixes three jobs and makes each worse. Splitting the work across focused specialists gives cleaner outlines, clearer notes, and better review questions — and it can all run on models hosted on your own machine.

Approach — One job per agent

A multi-agent system that turns any topic into a beginner-friendly study guide — a three-part outline, concise notes, and review questions — saved as clean Markdown, using models hosted locally on Ollama, LM Studio, or any OpenAI-compatible server.

Key Highlights

• Three specialist agents in a sequential pipeline: a planner writes the outline, a teacher turns it into notes, and a quiz writer creates review questions
• Plain Python + LangChain version where a controller function passes each agent's output to the next, timing every step
• LangGraph version of the same flow, with each specialist as a node, shared state, and edges START → planner → teacher → quiz → END
• Runs fully on local models through Ollama, LM Studio, or any OpenAI-compatible server such as vLLM or llama.cpp
• Documents common multi-agent patterns: parallel specialists, orchestrator–subagent, supervisor/router, human-in-the-loop, and review loops

The Problem

Placing emergency shelters optimally is an NP-hard problem — brute force is intractable at scale. This project uses evolutionary computation to find near-optimal shelter placements that maximize population coverage while respecting blast zone safety constraints and infrastructure accessibility.

Approach — Using evolutionary computation for real-world facility placement

A Genetic Algorithm for the Uncapacitated Facility Location Problem (UFLP) that identifies optimal nuclear shelter locations across ~30,000 US zip codes, maximizing population coverage while enforcing a 15-mile blast zone exclusion radius around urban targets.

Key Highlights

• Binary chromosome encoding with tournament selection, uniform crossover, and bit-flip mutation, evolving candidate solutions toward high-fitness placements
• Multi-objective fitness: population coverage within a serviceable radius, strategic safety (distance from nuclear targets), and infrastructure accessibility (road networks and power grid)
• Geospatial data from the US Census Bureau (zip code populations), nuclear target databases, and OpenStreetMap road networks
• Spatial joins, distance calculations, and exclusion zone masking with GeoPandas, Shapely, and OSMnx
• Benchmarked against a greedy baseline heuristic, with the GA better balancing competing objectives across large candidate sets

The Problem

Educational institutions need flexible LMS platforms that handle distinct user workflows — admins managing users, professors building courses, and students consuming content. Kambaz demonstrates end-to-end web development with complex authorization logic and a split deployment architecture.

Approach — Designing for three different user types

A full-stack learning management system similar to Canvas, with a React frontend, a Node.js backend, MongoDB for persistent storage, and a RESTful API.

Key Highlights

• Role-based access control with separate Admin, Professor, and Student views
• Course management, assignment submission, quizzes, exams, and grading workflows
• RESTful API design with a Node.js backend and MongoDB data layer
• Split deployment: frontend on Vercel and backend on Render, with CORS policies and production build pipelines