Uzair Arif
Building and researching multi-agent AI systems.
About Me
I build and research multi-agent AI systems where independent language models cooperate, reason, and make decisions together. My work focuses on solving the hardest failure modes in agentic systems: how groups form sustainable rules, how shared memory avoids contamination, and how models handle strategic risk.
In ELICIT, I designed a multi-agent simulation framework that tests how LLM agents negotiate climate treaties and self-govern under risk. Agents choose whether to contribute to shared pools, vote to rewrite constitutional rules, and manage an empirical Loss and Damage Fund. In SEAM, I study agentic memory architecture to stop memory collapse and information poisoning. By comparing update policies across network topologies, I showed how structured curation protects agent stability over hundreds of rounds.
Beyond research, I build applied ML systems that deploy to production. My background in Economics and Mathematics at IBA Karachi gives me a strong quantitative foundation for building real-time forecasting pipelines, hybrid retrieval engines, and scalable MLOps architectures.
Technical Skills
- Languages: Python, SQL
- ML / DL: Scikit-learn, PyTorch, XGBoost, CatBoost
- Time Series: Prophet, TCN, Darts, Recursive Forecasting
- NLP / LLM: SciBERT, LangChain, RAG Systems, Multi-Agent Simulation
- Retrieval: FAISS, BM25, Reciprocal Rank Fusion, Groq (Llama 3)
- Optimisation: Optuna, Feature Engineering, 5-Fold CV
- Data & Tools: Pandas, NumPy, Jupyter, Git, QGIS, ZeroMQ
- MLOps: MLflow, FastAPI, Docker, ZeroMQ, DVC
- Platforms: Streamlit, Kaggle, Google Colab
Featured Projects
ELICIT: Emergent LLM Institutions for Climate Treaties
Tags: Multi-Agent Systems, LLM, Game Theory, Climate Risk, Governance, Ollama
This research project tests whether LLM agents can build and follow their own rules to solve climate problems. Independent agents choose whether to contribute resources, reward or punish peers, and vote to update their governing policies across repeated rounds. The simulation includes climate shocks, a shared Loss and Damage Fund, and peer gossip. It shows how decentralized agent groups learn to cooperate under risk.
SEAM: Shared Evolving Agent Memory
Tags: Multi-Agent Systems, LLM Agents, Agentic Memory, AI Safety, Network Topologies, Ollama
This project studies what happens when self-evolving agents share memory with each other. It measures how quickly bad experiences or false lessons spread through a population across different network topologies. The benchmark compares naive overwrite, raw history logs, and structured incremental updates on deterministic resource tasks. Results show that structured curation with ring communication prevents performance collapse while isolating errors.
ECG5000 Anomaly Detection with TCN
Tags: Deep Learning, TCN, PyTorch, Self-Supervised Learning, SimCLR, Anomaly Detection, Time Series
This project builds a PyTorch pipeline to identify irregular heartbeats in ECG time series data. It uses Temporal Convolutional Networks across supervised, self-supervised, and contrastive learning setups. The system automatically calibrates detection thresholds to flag abnormal signals with high precision. It also outputs clear visual reports including reconstruction curves and confusion matrices.
Taxi Demand MLOps Pipeline
Tags: MLOps, XGBoost, Optuna, MLflow, FastAPI, SQLite, Docker, Drift Monitoring
This is a complete MLOps system that predicts taxi demand across New York City in real time. It cleans raw trip records, builds rolling time-window features, and trains an XGBoost model tuned with Optuna. Predictions run through a FastAPI service connected to a local feature store and containerized with Docker. It also tracks data drift continuously using statistical tests and includes 36 automated unit tests.
IBA RAG Chatbot
Tags: NLP, RAG, Hybrid Search, FAISS, BM25, LangChain, Groq, Streamlit
This is an internal assistant that answers student and faculty questions using university documents and web pages. It pairs FAISS vector search with BM25 keyword search through reciprocal rank fusion to find exact information. Answers are generated by Llama 3 via Groq with clear source citations for every claim. Users interact with the system through a responsive Streamlit interface.
Prophet Daily Revenue Forecasting
Tags: Time Series, Prophet, Optuna, Lunar Calendar, Revenue Forecasting
This tool forecasts daily business revenue while accounting for complex seasonal trends. It integrates Facebook Prophet with custom calendars for national holidays and Islamic lunar events. Model parameters are tuned automatically using Optuna to minimize prediction error. Forecast tables and trend charts are exported directly into Excel spreadsheets for operational planning.
Impostor Hunt: Detecting Hallucinated LLM Outputs
Tags: AI Security, NLP, CatBoost, SciBERT, GPT-2 Perplexity, Kaggle, Optuna
This project detects false and data-poisoned text generated by language models. It extracts text features using SciBERT sentence embeddings, GPT-2 perplexity scores, and vocabulary similarity metrics. A CatBoost classifier trained on these features achieved a top competition score of 0.9066. The full pipeline was validated using 5-fold cross-validation with Optuna tuning.
Inflation Forecast ML
Tags: Forecasting, Ensemble ML, Random Forest, XGBoost, Recursive Forecasting
This machine learning model predicts monthly national inflation rates using public economic indicators. It combines Random Forest and XGBoost into a weighted ensemble for consistent predictions. The model uses recursive multi-step forecasting and automated feature selection to model economic momentum. It delivered top-tier benchmark accuracy in the university forecasting challenge.
LSTM AAPL Stock Prediction
Tags: Deep Learning, LSTM, Time Series, yfinance, Keras
This project uses recurrent neural networks to study stock price patterns. An LSTM network processes historical trading data fetched directly from Yahoo Finance. The pipeline normalizes sequential prices and evaluates performance using RMSE and R-squared metrics. It generates visual comparison plots showing predicted price curves against actual market history.
Bitcoin Sentiment Prediction
Tags: Sentiment Analysis, XGBoost, Random Forest, Wikipedia, yfinance, Cryptocurrency
This project predicts next-day Bitcoin price movements by combining price data with public interest indicators. It extracts edit activity and traffic patterns from Wikipedia articles as a proxy for market sentiment. The combined dataset trains XGBoost and Random Forest classifiers to determine price direction. This approach helps evaluate how retail attention influences short-term cryptocurrency volatility.
Work Experience
Research Intern
National Center of Artificial Intelligence (NCAI) · Islamabad, Pakistan (On-site)
- Conducted literature reviews on Multi-Agent Reinforcement Learning (MARL), public goods games, endogenous institutional design, and Ostrom's governance principles.
- Drafted a research proposal titled "Emergent Institutional Design in Public Goods Games: A Multi-Agent Reinforcement Learning Approach" (IPPO architecture, voting systems), laying foundations for the eventual FYP (ELICIT).
Junior Data Analyst
Ali Khan Consultants · Karachi, Pakistan (Hybrid)
- Analyzed supply chain and sales pipeline data to identify process inefficiencies, uncovering a 32% (60,000 kg) forecast discrepancy.
- Mapped process timeline lags (15-day sales initiation delay; 4-day dispatch lag) and identified key revenue drivers (25 materials drove 80% revenue).
Revenue Forecasting Intern
Federal Board of Revenue (FBR) · Islamabad, Pakistan (Hybrid)
- Developed Prophet daily tax revenue forecasting models for Income Tax, Sales Tax, Federal Excise, Customs Duty, and Gross Collection.
- Implemented Optuna hyperparameter tuning, managed Pakistan-specific holiday/fiscal deadline seasonality, and cleaned outlier datasets.
Artificial Intelligence Intern
National Aerospace Science & Technology Park (NASTP) · Rawalpindi, Pakistan (On-site)
- Developed low-latency ZeroMQ real-time data streaming pipelines and live geospatial QGIS maps.
- Built CNN and SVM baseline models (improving test accuracy by 10%) and created an ECG time-series anomaly detection system (95% detection rate).
Get In Touch
Available for worldwide remote roles, AI research collaborations, and production ML opportunities.