Soroush Vahidi

LLM-Serving Scheduler Portfolios

Reproducible evaluation infrastructure for LLM-inference serving schedulers, spanning 120 workload windows, 13 scheduler policies, and 6 calibrated operating regions (9,360 configurations, 18,720 GPU-calibrated simulator executions), with a 240-run real-vLLM/GPU benchmark campaign on SLURM used to check simulator findings against instrumented hardware runs on a selected case. Per-scenario policy selection improved on the best single fixed policy; the project reports this result together with where lighter-weight online selectors fell short, rather than only the positive headline number. This is selected-case hardware validation, not a claim of broad multi-cluster validation.

Budget-Aware LLM Inference & Cost-Aware Reasoning Routing

Two related lines of work on spending a fixed inference budget well: (1) analyzing cost/latency/accuracy tradeoffs for closed-API LLM inference (OpenAI, Gemini) under matched logical-call budgets across GSM8K, MATH-500, GPQA-Diamond, and StrategyQA; and (2) an adaptive, cost-aware routing policy that decides when a cheap LLM answer should be revised versus accepted, evaluated under matched-budget conditions.

Retrieval-Assisted Natural-Language Optimization

A system that maps natural-language descriptions of combinatorial optimization problems to structured LP/ILP formulations, using retrieval-conditioned schema selection, deterministic grounding, and solver-backed validation (SciPy/HiGHS). Evaluated on a 331-query corpus together with the public NLP4LP benchmark and the OptMath external validation domain.

Learning-Augmented Cache Eviction

A trace-driven cache-eviction simulator studying whether predicting finite-horizon downstream miss costs at the candidate level can improve online cache replacement, benchmarked against LRU, SIEVE, FIFO-Reinsertion, and learned baselines (LRB, 3L-Cache, CACHEUS, HALP) across seven workload families. Under the corrected, matched-comparison evaluation protocol, the proposed predictor does not outperform the strongest baselines — the manuscript reports this as a controlled negative/mixed result rather than a claim of superiority. The accompanying LAFC-Evict dataset (v0.3, 22.4M rows of counterfactual eviction-decision supervision) is published on Hugging Face as a public artifact.

Training-Free Ranking & Weighted Feedback Arc Set

Training-free ranking from pairwise comparisons via weighted feedback-arc-set (FAS) and acyclic graph construction (accepted, The Journal of Supercomputing, 2026 — see Publications). Related large-scale-graph work includes an incumbent-protected SCC-neighborhood search and a separate interval-based local-refinement heuristic for the weighted FAS problem, applied to the FlyWire connectome (136,648 neurons, 5.6M+ weighted synapses), reporting the highest forward-weight score published to date on that benchmark.

NJIT Thesis Auditor

A free, local-first tool that audits NJIT thesis/dissertation formatting against Graduate Studies guidelines from a compiled PDF, TeX source tree, or DOCX comments — no external API calls for document analysis. Implements 80+ heuristic checks (TeX hygiene; PDF margins, fonts, spacing, headings, TOC/LOT/LOF structure, page numbering, front matter, and references) behind both a web UI and a CLI, with SSRF-guarded URL fetching, privacy-safe report exports, and a human-feedback workflow for reviewing PDF/DOCX annotations. Backed by 3,192 passing automated tests and green multi-version CI (Python 3.10–3.12), with clean mypy and ruff checks across 113 source files.

Repository is currently private; available on request.

Software Developer Intern — Canfield Scientific Inc., Parsippany, NJ

Summer 2025 (Jun–Aug 2025)

Contributed to a production C++/Qt dermatology imaging platform. Implemented modules for multi-camera image capture, annotation, lesion tracking, and report generation within an established production codebase; debugged real-time acquisition workflows and image-processing pipelines under production and memory constraints.

Graduate Research Assistant — NJIT

Jan 2022 – Mar 2023

Researched algorithmic and optimization-based approaches for computational problems, including parallel LCS algorithms in Chapel (IEEE HPEC 2023), energy-harvesting-aware real-time scheduling (Algorithmica 2025), and local-search methods for connected maximum cuts (ESA 2023).

Graduate Teaching Assistant — Department of Computer Science, NJIT

Apr 2023 – Present

Supported CS 610 Data Structures and Algorithms (multiple semesters, 2023–2026), DS 677 Deep Learning (Fall 2025), CS 435 Advanced Data Structures (Fall 2025), and CS 388 / CodePath Intermediate Android Development (Spring 2026). Duties included grading, student mentoring, exam support, review of GitHub-based lab and project submissions, repository-access troubleshooting, CodePath/Canvas grading workflows, and building Python-based grading/course automation tools.