Berkeley, CA · open to opportunities

Parham builds systems for autonomy

EECS & MEng at UC Berkeley. Research on safe imitation learning, multi-agent coordination, and LLM reasoning.

↓ Resume Projects Publications About GitHub ↗
Portrait of Parham Sharafoleslami
Now Decentralized fleet coordination for airport ground ops · MEng thesis on safe imitation learning for race-car controllers

Selected work

all 22 →
Robotics · 2025

Nonlinear MPC for Autonomous Racing

Progress-maximizing NMPC for a full-scale Dallara AV-24 competing in the Indy Autonomous Challenge. Two-timescale architecture: a full-lap minimum-time NLP solved offline with a double-track Pacejka vehicle model produces the Track Trajectory Library, while an NMPC at 100 Hz in a Frenet frame executes it with RTI. Validated in AWSIM (best lap 117.6 s, 207 km/h max) and on the physical car.

PythonC++17CasADiacadosIPOPTHPIPM
Read case study →
Robotics · 7 min read

MPC for UR7e Robotic Arm: Warehouse Sorting

Constrained joint-space Model Predictive Control for a UR7e manipulator. An RGB-D perception pipeline localizes objects and obstacles; a receding-horizon CasADi/IPOPT solver generates collision-free joint trajectories at 12 Hz, validated in MuJoCo and deployed on real hardware.

PythonMuJoCoCasADiROS2 +1
Robotics · 8 min read

Decentralized Fleet Coordination for Airport Ground Operations

Multi-agent coverage control for airport ground vehicles using Buffered Voronoi Cells, Lloyd's algorithm, and game-theoretic demand response. Zero collisions across all test scenarios.

PythonMulti-Agent SystemsOptimal ControlGame Theory
Machine Learning · 14 min read

Multi-Agent RL for Autonomous Driving in Waymax

PPO agent trained in Google's JAX-based Waymax simulator on Waymo Open Dataset traffic. Began as speed tracking on log-replay traffic and ended as adaptive cruise control evaluated on a five-scenario stress suite: stalled vehicle, slow lead, emergency braking, stop-and-go, and cut-in. Four pass with zero contact; the fifth exposes a measurable generalization gap. Fourteen training runs, ten documented failure modes.

PythonJAXFlaxOptax +2
Machine Learning · 3 min read

Deep Learning from Scratch

The full arc: backpropagation in raw NumPy, CNNs with BatchNorm and Dropout trained to 78% on CIFAR-10, multi-head self-attention for text summarization, and a Masked Autoencoder that reconstructs images from 25% of their patches, then transfers those features to downstream tasks.

PythonNumPyPyTorchCIFAR-10
Machine Learning · 4 min read

Classical ML from Scratch

Two learning paradigms built from NumPy up: tree-based spam classification (decision tree, Random Forest, AdaBoost) and SVD/ALS matrix factorization for movie recommendations. No frameworks; matched scikit-learn on both.

PythonNumPySciPyMachine Learning
Natural Language Processing · 5 min read

Modern NLP: From Statistical MT to Multimodal Foundation Models

Four paradigm shifts in one semester: IBM Model 1, attention-based NMT, transformer parsing, LLM fine-tuning, and CLIP multimodal retrieval with pragmatic reasoning. Each technique subsumes and extends the previous.

PythonPyTorchTransformersCLIP +1

Selected publications

all →
Preprint · 2026 · advisor

Visualizing and Benchmarking LLM Factual Hallucination Tendencies via Internal State Analysis and Clustering

Nathan Mao , Varun Kaushik , Shreya Shivkumar , Parham Sharafoleslami , Kevin Zhu , Sunishchal Dev

FalseCite, a curated dataset of 82k false claims paired with fabricated citations, reveals that LLMs hallucinate more readily when misleading references are present, especially in smaller models like GPT-4o-mini. Hidden-state clustering exposes a distinctive 'horn-like' geometry across hallucinating and non-hallucinating activations.

arXiv ↗ PDF ↗ #hallucination#benchmark#interpretability
Preprint · 2025

COMPASS: Context-Modulated PID Attention Steering System for Hallucination Mitigation

Kenji Sahay , Snigdha Pandya , Rohan Nagale , Anna Lin , Shikhar Shiromani , Parham Sharaf , Kevin Zhu , Sunishchal Dev

A decoding-time intervention that dynamically steers attention toward retrieved context using a PID controller driven by a per-head Context Reliance Score. No retraining, no multi-pass decoding: just interpretable, single-stream control of evidence grounding. Reduces hallucinations by 2.8–5.8% absolute across HotpotQA, XSum, HaluEval, and RAGTruth.

arXiv ↗ PDF ↗ #hallucination#decoding#attention#interpretability
NeurIPS 2025 · 2025 · advisor

Optimizing Chain-of-Thought Confidence via Topological and Dirichlet Risk Analysis

Abhishek More , Anthony Zhang , Nicole Bonilla , Ashvik Vivekan , Kevin Zhu , Parham Sharafoleslami , Maheep Chaudhary

EDTR estimates LLM confidence by treating each chain-of-thought as a vector in semantic space and analyzing the geometry of the reasoning distribution. Combined with Dirichlet-based uncertainty quantification, it achieves 41% better calibration than competing methods and perfect accuracy on AIME.

arXiv ↗ PDF ↗ #calibration#chain-of-thought#uncertainty

Working with

Python 17 NumPy 9 PyTorch 4 Computer Vision 4 CasADi 2 ROS2 2 Optimal Control 2 Game Theory 2 SciPy 2 scikit-image 2 RISC-V 2 Information Theory 2 Matplotlib 2 Java 2 C++17 1 acados 1