Murtaza Nikzad

Now

• Forest Systems

  Since July 2026

  Software developer assistant, remote from San Francisco. I build AI-assisted testing infrastructure for financial software: LLM-generated synthetic test data, semantic assertion pipelines across Python and .NET services, and document-extraction evaluation for private equity statements.

• Horizon3.ai

  May – June 2026

  Machine learning research intern. I build LLM agents for autonomous offensive security: reinforcement learning, agent harnesses, and multi-turn tool use on the NodeZero platform.

• Davidson College

  Since August 2024

  Independent research with Raghuram Ramanujan. A single model proposes problems and solves them, trained with GRPO on verifiable rewards: self-play curriculum generation for multi-turn, multi-tool agentic RL. The work extends my DPO paper.

• Cleveland Clinic

  May 2025 – May 2026

  Machine learning research in clinical prediction. I built TremorAnalyzer, a signal processing pipeline used in clinical trials, and deep learning models that predict HIFU lesion outcomes at 85% accuracy across 100+ patients. The software informs neurosurgical treatment planning.

The paper

Forward versus Backward: Comparing Reasoning Objectives in Direct Preference Optimization

With Raghuram Ramanujan · arXiv, January 2026

I fine-tuned Llama 3.1 8B-Instruct with DPO on GSM8K under two objectives. Forward learns to generate the chain of thought. Backward learns to verify: given a problem and a candidate answer, the model judges pass or fail.

Preference optimization makes models more confident and less willing to admit their own errors, even as accuracy improves.

All numbers are percentages on GSM8K. Acknowledgement is how often the model flags its own wrong answers.

arXiv · PDF

Earlier: Transfer Learning with PointNet for Time Projection Chambers, a poster with R. Doctor, M. P. Kuchera, R. Ramanujan, D. Bazin, and Y. Ayyad at the Davidson Summer Research Symposium.

Selected work

• Hack@Davidson

  President

  I run Davidson’s hackathon. The 2026 edition drew 229 registrations from 35+ universities.

  Site

• Gaudi

  June 2026

  Draw a 2D floor plan in the browser and it becomes a navigable 3D house. Room detection by planar face traversal, CSG-free openings, and AI furnishing that generates and then verifies against a deterministic geometry validator.

  Live · Code

• Rover VLA

  HUD × YC RSI Hackathon

  A fine-tuned vision-language-action model drives a rover to a person. Trained on MuJoCo sim demonstrations; 95% success on held-out layouts against a 0% random baseline.

  Code

• Artki

  Cal Hacks Chroma Prize

  AI interior design. CLIP embeddings learn what you like, Claude generates Three.js rooms, ChromaDB runs the furniture search.

  Code

• Open Context

  HackHarvard Best AI Application

  AI study platform: chat, OCR lecture transcription, and quiz generation over a Pinecone RAG pipeline.

  Site · Code

• Paddington

  Spring 2025

  Real-time voice AI on WebSockets, Whisper, and Llama 3.1. Deterministic sub-1.5s latency; streaming optimizations cut perceived latency by 60%.

  Code

Five hackathon wins overall: Cal Hacks, HackHarvard, VandyHacks, HackNC, and HopHacks.

Elsewhere

Résumé · GitHub · Twitter · LinkedIn · munikzad@davidson.edu