CV

Sem-DPO: Mitigating Semantic Inconsistency in Preference Optimization

Advisors: Dr. Ali Anwar, University of Minnesota · Dr. Ahmad Faraz Khan, Google

Published at ACL 2026

• Developed a weighted Direct Preference Optimization approach that reduces semantic drift during automated prompt optimization and established a theoretical bound on drift in preference-tuned prompt generators.
• Improved CLIP similarity by 8–12% and human-preference scores by 5–9% over standard DPO across three benchmarks and three language-model backbones; preferred in 82.2% of pairwise human evaluations.

LADs: Leveraging LLMs for AI-Driven DevOps

Advisors: Dr. Ali Anwar, University of Minnesota · Dr. Ali R. Butt, Virginia Tech

First author. Submission at MLSys 2027

• Co-developed a six-agent framework for configuration, deployment, monitoring, and automated repair; achieved 68% deployment success across 180 Dask, Redis, and Ray tasks and repaired 94% of runtime failures within three iterations.
• Adapted two existing systems as baselines and substantially redesigned the evaluation and technical presentation to better establish LADs’ gains over prior approaches.

SkillGraph: Reusable Reasoning Skills from LLM Agent Traces

Advisors: Dr. Tianyin Xu, University of Illinois Urbana-Champaign · Dr. Ali Anwar, University of Minnesota

First author. Submission at MLSys 2027

• Leading development of a GraphRAG-based skill graph for extracting and reusing diagnostic reasoning from agent trajectories; built a database of reasoning traces for skill extraction and retrieval on new system incidents.
• Identified shortcut behavior that corrupts learned skills and designed a process-level critic that evaluates reasoning quality rather than final outcomes alone.

Agent Memory: Storage and Retrieval of LLM Reasoning Traces at Scale

Advisor: Dr. Ali Anwar, University of Minnesota

Submission at MLSys 2027

• Analyzed 535 queries across seven models, finding reasoning traces from under 1 KB to over 71 KB per query; developing co-location and indexing strategies for highly variable semantic reasoning workloads.

Federated Skill Sharing for LLM Agents

Advisors: Dr. Ali R. Butt, Virginia Tech · Dr. Ahmad Faraz Khan, Google

Submission at MLSys 2027

• Developing a federated framework where clients share reusable skill templates and generalized reasoning patterns through a global indexed library while keeping private context and final adaptation local.

Summer Intern

Systems Limited

Jun 2025 – Jul 2025

• Built an ERP workflow automating procurement, project tracking, and reporting for a construction company using Microsoft Dynamics.

B.S.

Lahore University of Management Sciences

Computer Science

• Relevant Coursework: Data Structures & Algorithms, Operating Systems, AI/ML, Generative AI, Advanced GenAI, AI on edge devices, Network Centric Computing, Topics in LLMs

Sem-DPO: Mitigating Semantic Inconsistency in Preference Optimization for Prompt Engineering

Findings of ACL 2026

A weighted Direct Preference Optimization approach that reduces semantic drift during automated prompt optimization, with 8–12% higher CLIP similarity and 5–9% higher human-preference scores over standard DPO.

LADs: Leveraging LLMs for AI-Driven DevOps

MLSys 2027 (Under Submission)

A six-agent framework for configuration, deployment, monitoring, and automated repair. First author. 68% deployment success across 180 Dask, Redis, and Ray tasks, and 94% of runtime failures repaired within three iterations.

SkillGraph: Reusable Reasoning Skills from LLM Agent Traces

MLSys 2027 (Under Submission)

A GraphRAG-based skill graph for extracting and reusing diagnostic reasoning from agent trajectories, with a process-level critic for reasoning quality. First author.

Agent Memory: Storage and Retrieval of LLM Reasoning Traces at Scale

MLSys 2027 (Under Submission)

Analysis of 535 queries across seven models, with reasoning traces from under 1 KB to over 71 KB per query, toward co-location and indexing for variable semantic workloads.

Federated Skill Sharing for LLM Agents

MLSys 2027 (Under Submission)

A federated framework for sharing reusable skill templates and generalized reasoning patterns while keeping private context and final adaptation local.