Manuscript in preparation2026
Directional Feedback Guided by Failure Modes for Inference-Efficient LLM Reasoning in Telecom Mathematics
M. A. Shakeel, N. U. Hassan
Available for PhD & freelance engagements
Muhammad Ahsan Shakeel
AI Researcher & Machine Learning Engineer.
MSc AI at LUMS researching inference-efficient LLM reasoning. I ship production AI systems — agentic pipelines, RAG, MLOps — and publish peer-reviewed work on the side.
Lahore, Pakistan·LUMS · MSc AI '26·IEEE PIMRC 2026
Current research
Failure-Mode-Guided Directional Feedback for LLM reasoning
About
Research-grade rigour, production-grade delivery.
I'm an MSc Artificial Intelligence candidate at the Lahore University of Management Sciences, working with Dr. Naveed-ul-Hassan on inference-efficient LLM reasoning. My thesis introduces FMGDF — a three-pass agentic pipeline that uses failure-mode taxonomies to maximise accuracy under tight call-count and token budgets.
Alongside research I teach 300+ undergraduates at the University of Lahore and ship LLM systems for industry — RAG, agentic workflows, MLOps pipelines on AWS Bedrock and Kubernetes. I care about the boring parts: reproducibility, evaluation harnesses, cost per call, and graceful failure.
"Most LLM gains come from disciplined feedback loops, not bigger models."
For PhD committees
Research & publications
Focused on LLM reasoning, agentic systems, and the physics of learning under constraint. Actively seeking PhD positions starting Fall 2026.
IEEE PIMRC2026
Intent-Driven LLM-Based Two-Level Control in AI-Native 6G Networks
Z. Hussain*, M. A. Shakeel*, N. U. Hassan, H. Chen, C. Yuen
Submitted — IEEE Proceedings2025
Noise Resilience of Quantum Support Vector Machines: Feature Map Architectures for NISQ Deployment
M. A. Shakeel et al.
Research experience
Sept 2024 — Present
LUMS, EE Department
Graduate Research Assistant — LLM Reasoning & Inference Efficiency
- Designed FMGDF, a three-pass agentic inference pipeline (critic LLM, failure-mode taxonomy, diversity prefix) maximising accuracy under joint call-count and token budgets.
- Benchmarked AWS Bedrock (Claude 3 Sonnet, Llama 3.1-70B) against self-hosted OSS models behind a unified Python interface — model swaps via config change only.
- Tracked experiments with MLflow and built an automated evaluation harness — cut manual reporting time by 70%.
2025
LUMS, EE Department (PIMRC 2026)
Research Contributor — LLM-Driven 6G Network Control
- Co-designed intent-driven two-level control with an LLM upper layer refining Lyapunov parameters via structured feedback.
- Validated across GPT-4.5, GPT-OSS-120B, and Llama-3.3-70B with full integration tests on every API contract.
2024 — 2025
LUMS, Physics Department
Research Contributor — Quantum Machine Learning
- Ran 52 controlled experiments across four QSVM feature-map architectures and three noise types.
- Amplitude-inspired encoding sustained 100% accuracy up to noise level 0.10 — a 10× gain over the next-best approach.
For founders & teams
Freelance services
Engagements range from rapid LLM prototypes to production MLOps overhauls. Available for contract, retainer, or advisory.
01
LLM Agents & RAG Systems
Multi-turn agentic workflows with LangChain, Bedrock Agents, FAISS / Chroma vector stores, tool-use, and Knowledge Bases.
- Production agent pipelines
- Vector store + ingestion
- Tool & function-calling design
02
MLOps & Deployment
From notebook to production: Docker, Kubernetes, GitHub Actions CI/CD, MLflow tracking, DVC versioning, AWS deployment.
- Reproducible training pipelines
- CI/CD for models & prompts
- Monitoring & rollback strategy
03
Fine-Tuning & Model Evaluation
PyTorch fine-tuning, structured A/B experiments, bias audits, calibration analysis, and responsible-AI reporting.
- Domain-adapted models
- Evaluation harness
- Compliance documentation
04
FastAPI Microservices
Async REST endpoints, OpenAPI schemas, and provider-agnostic backends so model swaps require only a config change.
- Production API service
- OpenAPI contract
- Container + deployment
05
Computer Vision & Classical ML
Image classification, tabular ML (XGBoost, ensembles), imbalanced-data techniques, ONNX export, REST serving.
- Trained & exported model
- Inference endpoint
- Validation report
06
Data Pipelines & ETL
Airflow DAGs, Spark transformations, Hadoop/Hive batch processing, Dask for distributed workloads.
- Scheduled DAGs
- Distributed ETL
- Data quality checks
07
Research Collaboration
Co-author on ML / LLM / quantum-ML papers — problem framing, experimental design, ablations, LaTeX manuscript prep, and rebuttal support for IEEE / NeurIPS-tier venues.
- Co-authored manuscript
- Reproducible experiment repo
- Reviewer-response pack
08
Research Assistant (Remote)
Literature reviews, dataset curation, baseline reproductions, and end-to-end experiment running for PhD students, labs, and independent researchers.
- Annotated lit review
- Curated dataset + scripts
- Weekly progress reports
09
Virtual Assistant — Technical
Long-term remote support for founders and academics: inbox triage, calendar and travel, manuscript formatting, slide decks, data entry, and light automation in Python / Zapier.
- Inbox & calendar SOPs
- Polished decks & docs
- Automation scripts
Selected work
Projects worth talking about
LLM Agent Orchestration
Multi-agent workflow with LangChain tool use, REST calls for real-time data, deployed as an AWS microservice.
LangChainAWSAgents
Solar Inverter Control Interface
Designed and built a touch-friendly monitoring & control UI for a solar inverter — real-time telemetry (V, I, kWh, battery SOC), fault alerts, and Modbus/MQTT backend integration.
Embedded UIMQTTModbus
Naruto Jutsu Recognition (CV)
Real-time hand-sign / mudra classifier using MediaPipe landmarks + a lightweight CNN over OpenCV — maps sequences to jutsu and overlays anime-style visual effects on the webcam feed.
OpenCVMediaPipePyTorch
Melanoma Detection
Binary classifier on HAM10000 with Dask distributed computing. 92.81% accuracy. ONNX export + REST.
PyTorchDaskONNX
Real Estate Voice Agent
End-to-end agentic assistant with RAG over FAISS and Chroma for property search and legal Q&A.
RAGFAISSFastAPI
Bankruptcy Prediction
LR, RF, XGBoost, GBM compared with SMOTE on imbalanced finance data. Full lifecycle in MLflow.
XGBoostMLflowSMOTE
Big Data Pipeline
ETL on HDFS + Hive across a multi-node Hadoop cluster, extended with Apache Spark for analytics.
HadoopHiveSpark
Trajectory
Experience
Jun 2024 — Present
University of Lahore
Junior Lecturer, CS & IT
- Teach Programming Fundamentals and OOP to 300+ undergraduates per semester.
- Designed a hands-on lab on fine-tuning small LMs and deploying via FastAPI to cloud.
- Introduced an applied AI-ethics module on bias auditing and responsible generative AI.
Jul 2024 — Feb 2026
Phoenux Design (Remote)
AI & Web Projects Associate
- Built agentic workflows on LangChain + Bedrock; containerised with Docker, orchestrated on Kubernetes — cut p95 latency by 35%.
- Implemented DVC + MLflow so every deployed model was reproducible from a single config.
- GitHub Actions CI/CD dropped deployment effort from hours to under 10 minutes.
2022 — 2024
Bugsfree Solutions · Phoenux Design
AI Research Intern / Web Designer
- Fine-tuned PyTorch text classifier, served via FastAPI, integrated into a client SaaS.
- End-to-end feature engineering on 500k+ rows — F1 improved 18% via SMOTE + RFECV.
- Airflow DAGs for nightly refresh and automated retraining; DVC for dataset versioning.
Toolkit
Technical stack
Generative AI & Agents
LLM fine-tuningRAGLangChainAWS BedrockBedrock AgentsFAISSChroma
ML & AI Frameworks
PyTorchTensorFlowScikit-learnXGBoostQiskitCNNsTransformersONNX
MLOps & Deployment
MLflowDVCAirflowFastAPIDockerKubernetesGitHub ActionsAWS EC2 / ELB / SageMaker
Data & Distributed
PandasNumPyDaskHadoop / HDFSHiveApache SparkETL
Programming
PythonJavaC++SQLDartLaTeX
Get in touch
Let's talk — research or build.
Whether you're recruiting for a PhD cohort or need an AI engineer for a shipping deadline, I read every message.