PDF 中文

Kaleem Ullah Qasim

PhD in Artificial Intelligence | AI Researcher & LLM Engineer

Chengdu, China| +86-13111895637| kaleem@my.swjtu.edu.cn

LinkedIn | UpWork | GitHub | Google Scholar | HuggingFace

Research Areas

LLM & SLM Reasoning | Recursive Task Decomposition | Agentic Systems & Multi-Agent Coordination | RL Alignment (RLHF, DPO, GRPO, Verifiable Rewards) | Context Engineering

Technical Skills

Programming & Development: Python | TypeScript | JavaScript | SQL | FastAPI | Django | Flask | Git | REST APIs | GraphQL
LLM & Generative AI: HuggingFace (Transformers, TRL, PEFT, Datasets, Accelerate) | Fine-tuning (SFT, LoRA, QLoRA) | vLLM | DeepSpeed | FSDP | RAG Architecture | Context Engineering | In-context Learning
Agentic Systems: MCP (Model Context Protocol) | OpenAI Agents SDK | Pydantic AI | smolagents | Google ADK | LangGraph | LangChain | LlamaIndex | AutoGen | CrewAI | ReAct Pattern | Tool Use & Function Calling
Alignment & Agentic RL: RLHF (PPO, REINFORCE) | Preference Optimization (DPO, CPO, ORPO) | GRPO | Verifiable-Reward RL | TRL | verl | OpenRLHF | Reward Modeling | Multi-Agent RL
Machine Learning & Evaluation: PyTorch | scikit-learn | XGBoost | LightGBM | BERT | Transformers | NLP | Transfer Learning | lm-evaluation-harness | OpenCompass
MLOps & Cloud: AWS (SageMaker, Lambda, S3) | Huawei Ascend | Docker | Kubernetes | MLflow | Weights & Biases | Vector Databases (Pinecone, Weaviate, ChromaDB)

Professional Experience

Research Contract – Recursive Reasoning Agent System, Huawei Technologies (Lead Researcher)

2025 – Present
  • Won Huawei-commissioned research contract on the strength of RDoLT paper (JAIR Q1, 16 citations) as lead researcher, building a recursive decomposition agent framework for complex reasoning
  • Designed core recursive decomposition logic and memory modules, validated on AIME, U-MATH, and custom algorithm benchmarks – demonstrating measurable gains over pure CoT baselines
  • Evaluated heterogeneous large/small model combinations to optimize performance-cost tradeoffs; researched and selected RL training strategies for agent improvement
  • Building RL training loop on Huawei Ascend platform targeting top rankings on AIME/U-MATH benchmarks, with goal of CCF-A publication or patent

AI Engineer & LLM Specialist, Upwork (Freelance)

2023 – Present
  • Top Rated (top 10%) with 100% job success rate and 5-star ratings from 20+ clients across AI/ML projects
  • Built production RAG chatbots using LangChain/LlamaIndex/CrewAI with Pinecone and Weaviate vector databases, reducing task completion time by 20% with 95% accuracy on domain-specific queries
  • Optimized local LLMs (Llama 2/3, Mistral) using LoRA/QLoRA fine-tuning, improving task accuracy by 25% while ensuring GDPR compliance for privacy-focused enterprise deployments
  • Developed multi-agent AI systems using LangGraph and AutoGen for workflow orchestration, boosting semantic search accuracy by 35% and reducing API latency by 40%
  • Architected multimodal AI pipelines (Deepseek OC) for document analysis, processing 10K+ documents monthly with 92% accuracy using custom prompt engineering

Research Contractor – Traffic AI, University of Jeddah (Dr. Tariq Alsahfi)

2024 – Present
  • Co-authored 2 papers on LLM-based traffic analysis published in Alexandria Engineering Journal and arXiv, focusing on spatio-temporal reasoning and accident severity prediction using hybrid LLM architectures
  • Developed TraffiCoT-R framework combining Chain-of-Thought prompting with multi-agent coordination for traffic prediction, integrating GPT-4 with GIS data and graph neural networks (PyTorch Geometric)
  • Built SAFE (Semantic-Augmented Fusion Ensemble) hybrid system using Random Forest and Qwen3-4B LLM for accident severity classification, achieving 85.7% recall on severe accidents (from 0% with traditional ML) and 53.1% overall accuracy on naturally imbalanced data

Research Contractor – AI Security, Zhejiang University (Dr. Haitao Xu)

2022 – 2024
  • Co-authored paper on online affiliate marketing published in IEEE INFOCOM 2025, conducting empirical study of deceptive marketing practices and affiliate network analysis using NLP and network graph analysis techniques
  • Developed ADsFlow Chrome extension for detecting dynamic web advertisements, analyzing ad positioning and loading patterns on webpages using DOM analysis and computer vision (OpenCV) for real-time ad classification
  • Built webpage classification system using fine-tuned RoBERTa and BERT embeddings to identify webpage intent (cybersecurity threats, marketing, phishing), processing webpage code and content for multi-class classification (patent application filed)
  • Designed ML pipeline using XGBoost for deceptive ad detection, achieving 89% precision on 50K+ samples with automated feature extraction from URL structure, content, and affiliate link patterns

Data Scientist, Chengdu Ayurveda Biotechnology Co., Ltd

2020 – 2023
  • Contributed to company's achievement of #1 ranking on Alibaba marketplace through ensemble ML models (Random Forest, Gradient Boosting) for demand forecasting and dynamic pricing optimization, supporting 180% YoY revenue growth
  • Implemented predictive SEO strategy using time-series forecasting (ARIMA, Prophet) and NLP-based keyword analysis, achieving 95% increase in search appearances and 65% improvement in organic traffic over 12-month period
  • Built real-time analytics dashboard using Streamlit, Plotly, and PostgreSQL with automated ETL pipelines for sales, inventory, and customer data, improving data-driven decision response time by 60%

Publications

Selected Research Projects

RDoLT

JAIR (Q1) · 2025 · First Author
  • Recursive task decomposition with knowledge propagation, expanding reasoning into a verified sub-problem tree; significant gains over CoT/ToT on GSM8K, MATH, MMLU-STEM (16 citations; basis for Huawei research contract).

VERIFY-RL

arXiv 2026 · First Author
  • Process-level verifiable rewards combined with GRPO for end-to-end RL on 7B reasoning models over recursive decomposition trajectories; consistent gains over SFT and vanilla-GRPO on AIME and U-MATH.

CGAR – Curriculum-Guided Adaptive Recursion

arXiv 2025 · First Author
  • Curriculum-driven adaptive recursion-depth scheduling that adjusts inference steps per sample difficulty; accelerated training and lower inference cost while preserving accuracy.

Education

Ph.D. in Artificial Intelligence

2022 – 2026

Southwest Jiaotong University (SWJTU), School of Computing and Artificial Intelligence, Chengdu, China

Research: LLM/SLM Reasoning, Recursive Decomposition, Agentic Systems, Reinforcement Learning Alignment

Master in Computer Application Technology

2019 – 2022

Southwestern University of Finance and Economics (SWUFE), Chengdu, China

Focus: Natural Language Processing, Machine Learning, Deep Learning

Certifications

Professional Certifications: Google Prompting Essentials | Generative AI: Prompt Engineering | Generative AI & LLMs | Introduction to AI | Claude Code in Action | Reinforcement Fine-Tuning LLMs With GRPO | Post-training of LLMs

Languages & Communication: English (Fluent), Chinese (HSK 5 & HSKK Intermediate), Urdu (Native), Hindi (Fluent)