Saman Dehghan
Systems Software Engineer & Graduate Researcher
Bridging High-Performance Systems & AI for Software Engineering
I am a Systems Software Engineer and graduate researcher at the University of Illinois Urbana-Champaign (UIUC), working in the Intelligent-CAT Lab. With 5+ years of hands-on industry experience in low-latency C and C++, I specialize in constructing scalable, fault-tolerant distributed systems, network security appliances, real-time trading engines, and high-performance multithreaded data processors.
My current research sits at the intersection of Programming Languages (PL), Software Engineering (SE), and Large Language Models (LLMs). Specifically, I focus on:
- Automated C-to-Rust Translation: Eliminating unsafe memory bugs by translating legacy C systems into idiomatic, verifiable safe Rust (Rustine).
- Autonomous Programming Agents: Analyzing plan compliance, drift detection, and online corrective steering for SWE agents (ASE '26, LivePlan).
- Static Analysis & Compilers: Upstream LLVM/Clang profiling, coverage instrumentation (MC/DC), and kernel hypervisor internals (Xen).
-
High-Throughput Systems: SIMD vectorization, Linux
io_uringasync runtimes, kernel bypass (DPDK/Netmap), and lock-free concurrency.
Education
M.Sc. in Computer Science
University of Illinois Urbana-Champaign (UIUC)
Research with Intelligent-CAT Lab focusing on Large Language Models for Code, Program Analysis, and Software Engineering.
B.Sc. in Computer Engineering
Sharif University of Technology
Rigorous fundamentals in computer architecture, operating systems, algorithm design, and software engineering. GPA: 3.1 / 4.0.
Publications & Preprints
Peer-reviewed conference papers, benchmark studies, and preprints
From Plan to Action: How Well Do Agents Follow the Plan?
Investigates whether autonomous programming agents adhere to user/system generated plans during complex software development tasks. Evaluated over 21,120 SWE-agent trajectories across four state-of-the-art LLMs, eight plan variations, on SWE-bench Verified and SWE-bench Pro. Uncovers core dynamics of agent compliance, how structured phases influence issue resolution, and when reminders prevent plan deviation.
Online Monitoring and Corrective Steering of Programming Agents (LivePlan)
Introduces LivePlan, an execution framework that monitors programming agent trajectories in real time and applies corrective steering when agents drift or execute repetitive invalid actions. Separates deterministic rule-based triggers from LLM advisors, yielding up to 15.2% improvement (avg. 9.9%) in issue resolution on SWE-bench at a modest cost of $0.08 per instance.
Translating Large-Scale C Repositories to Idiomatic Rust (Rustine)
Presents Rustine, a fully automated repository-level pipeline that translates unsafe C codebases into idiomatic, memory-safe Rust. Evaluated on 23 C repositories up to 13,200 LOC, achieving 87% functional equivalence across 1,221,192 test assertions, drastically minimizing raw pointer usage, unsafe blocks, and compiler linter warnings compared to six baseline methods.
Assessing Code Reasoning in Large Language Models: A Literature Review of Benchmarks and Future Directions
Comprehensive survey of evaluation benchmarks and methodologies measuring how LLMs understand program logic, execution paths, and semantic invariants, identifying current blind spots in standard code benchmarks.
Research & Industry Experience
5+ years architecting and deploying production systems
Research Contributor
- Designing an automated LLM pipeline to translate unsafe C repositories into idiomatic, memory-safe Rust (Rustine).
- Building static analysis frameworks to detect unsafe pointer and concurrency patterns in C codebases and guide synthesis of safe Rust abstractions.
- Investigating plan-following dynamics and online monitoring for autonomous software development agents on SWE-bench.
Senior Software Engineer
- Architected and developed a real-time algorithmic trading platform executing strategies across crypto, forex, and commodity markets.
- Engineered high-throughput, low-latency FIX and FAST protocol server engines for rapid market data dissemination and order execution.
- Implemented quantitative option-pricing algorithms and a bidirectional MetaTrader 5 (MT5) integration bridge.
Software Architect
- Led the architectural design of a high-speed line-rate distributed network Data Loss Prevention (DLP) solution.
- Built distributed service discovery engines using Avahi & Poco and designed high-throughput Software-Defined Networking (SDN) dynamic routing tools.
- Mentored engineers, established code review benchmarks, and led cross-functional technical specifications.
Senior Software Engineer
- Refactored Redis into a multithreaded engine, optimized tail latency, and implemented reliable multicast cluster messaging.
- Constructed an ultra-low latency IPC framework combining shared memory buffers and lockless ring queues for cluster communication.
- Designed a high-throughput C++ concurrency library for streaming parallel data-processing pipelines.
C++ Developer
- Engineered a deep packet inspection (DPI) library for high-speed protocol classification across OSI network layers.
- Implemented high-performance pattern-matching engines implementing Aho–Corasick and RETE network algorithms.
Open Source & Key Systems Contributions
Upstream contributions to core systems, SIMD runtimes, and standalone projects
Xen Hypervisor
Added Clang 11–20 LLVM raw profile support (versions 5–10) and Modified Condition/Decision Coverage (MC/DC) profile bitmap sections for hypervisor test verification.
Highway & gemma.cpp
Accelerated LLM inference via Highway SIMD vectorization for RMSNorm and SquaredL2 in gemma.cpp. Implemented partial sort, nth-element, and heap selection in Highway.
Academic Service & Organization
Conference and workshop leadership
ReCode 2026 (1st Intl. Workshop on Code Translation, Transformation, & Modernization)
Role: Organizing Committee & Web Co-Chair
ReCode 2027 (2nd Intl. Workshop on Code Translation, Transformation, & Modernization)
Role: Organizing Committee & Web Co-Chair
Technical Skills & Expertise
Proficiencies across systems engineering, compilers, and modern AI/PL research