Arnav Verma
AI Engineering & Research
I work on language model safety and evaluations, representation learning and the systems that make it practical. Currently a Masters student at NYU Courant. Currently working on efficient training and evaluation of multimodal models.
๐ New York, NY
- Passed the CKAD exam โ Certified Kubernetes Application Developer
- Paper accepted to Findings of AACL-IJCNLP 2026
- Started the M.S. in Computer Science at NYU Courant
- Check Out Marigold - An App to help you study
Education

New York University, Courant Institute of Mathematical Sciences
Expected 05/28
M.S. Computer Science
Washington Square Campus, New York, NY
Affiliations
- Tech@NYU
- HackNYU
- IEEE Computer Society
Coursework
- Fundamental Algorithms - Yap CheeIn progress
- Machine Learning - Kyunghyun ChoIn progress
- Deep Learning - Yann LeCunnIn progress

University of North Texas
01/23 โ 05/26
B.S. Computer Science ยท Mathematics Minor
Dallas-Fort Worth, TX ยท GPA 3.70
Affiliations
- Responsible AI Lab
- UNT Computer Science Club
- UNT AI Research Program
- IEEE Computer Society
Awards & Honors
- President's List
- Dean's List
- Undergraduate Research Fellowship
- International Education Scholarship
- 1st Place โ HackSMU '24 & '25, HackUNT '23 & '24
- 2nd Place โ HackUTD '24 ยท Runner-up โ HackTX (UT Austin)
Coursework
- AlgorithmsA+
- Artificial IntelligenceA+
- Natural Language ProcessingA+
- Linear AlgebraA+
- Data StructuresA
- Operating SystemsA
- Database SystemsA
- Computer NetworksA
- Discrete MathematicsA
- Probability & StatisticsA
Research
Papers and preprints. Currently submitted, under review or published
Evaluating Multi-Dimensional Generalization of Large Language Models in Temporal Extraction Tasks
Iqbal, F., Dutt, R., Das, S., Verma, A., Ray Choudhury, S.
Findings of AACL-IJCNLP 2026. Accepted.
Do Large Language Models Reason Causally, or Just Pattern Match? A Multi-Strategy Evaluation
Verma, A., Iqbal, F., Das, S, Ray Choudhury, S.
Under review at A* conference
Projects
A few things I've built and written up.

Marigold - Turn Lecture Notes to Timed Quizzes and Flashcards
Upload a PDF of your notes, get flashcards generated from it, study them, then take a timed quiz on the same material and see exactly what you missed. Re-reading feels like studying; being asked and getting it wrong is what sticks.
- React
- FastAPI
- SQLite
- Gemini

Monte Carlo Options Pricer
A Monte Carlo engine for pricing European and American options. With variance reduction, Greeks, and alternative asset dynamics and a backtest that puts its deltas to work hedging a real position on ten years of S&P 500 data.
- Hedge Funds
- Trading
- Quantitative Research

Vectorized Portfolio Backtesting Engine
A daily-frequency engine for multi-asset equity strategies, built around the things that actually decide whether a backtest means anything: vectorised computation, explicit look-ahead handling, real transaction costs and honest out-of-sample validation. The headline result is negative, which is the point.
- Quantitative Research
- Quantitative Trading
- Python

Synapse - A workflow Automation Engine
Building Synapse: a workflow engine that survives crashes You draw a workflow as a graph in the browser, hit publish, and a fleet of workers runs it reliably. If a worker dies halfway through, the work still gets done.
- PyTorch
- TensorRT
- CUDA

Semantic Video Search Engine
Point it at a folder of videos and type "a red car driving at night". It returns the exact timestamps, ranked across every video you own, each one playable at the moment it matched โ searching the picture itself, not filenames or transcripts.
- PyTorch
- CLIP
- CUDA

Polish โ a Cursor-like AI Document Editor
Upload a document in any format, edit it inline with an AI assistant that sees the surrounding text, and export back out to PDF, DOCX or LaTeX. Version history means an aggressive rewrite is never a one-way door.
- Next.js
- TypeScript
- Express
- PostgreSQL
- Gemini
Writing
Posts on things I've worked through and wanted to explain properly.

Adapting LLaMA for NER Tasks
24 September 2025 ยท 7 min read
Customize your models using PEFT. Fine-tuning LLaMA 3.2 1B for token classification with LoRA adapters and 4-bit quantization, on a single GPU.
- Large Language Models
- LLM Finetuning
- NLP
- AI
- Llama 3

Language Model Distillation
18 March 2026 ยท 13 min read
Training a small student model to behave like a large teacher: what the soft probabilities actually carry, how to transfer attention patterns when the two models are different sizes, and how to tell whether it worked.
- Knowledge Distillation
- LLM
- Model Compression
- PyTorch

Kubernetes Pods Explained
2 August 2026 ยท 26 min read
What a Pod actually is, starting from the pause container nobody mentions: the lifecycle and its phases, imperative versus declarative creation, the manifest field by field, and the multi-container patterns worth knowing.
- Kubernetes
- Containers
- DevOps
- Infrastructure
Skills
Tools I reach for most often, shaded by how well I know them.
- Proficient
- Working knowledge
- Basic familiarity
Languages
- Python
- C/C++
- TypeScript
- JavaScript
- Rust
AI / ML
- PyTorch
- TensorFlow
- Keras
- Scikit-learn
- NumPy
- Pandas
- JAX
- OpenCV
- YOLO
- spaCy
Agentic AI & LLMs
- LangChain
- LlamaIndex
- HuggingFace
- Ollama
- Claude Code
- OpenClaw
- n8n
- Google AI Studio
Model Serving & Demos
- vLLM
- TensorRT
- Gradio
- Streamlit
- ONNX
Databases
- PostgreSQL
- MongoDB
- Pinecone
- Chroma
Backend
- FastAPI
- Flask
- Express
Frontend & Mobile
- React
- Next.js
- React Native
- Flutter
DevOps & Infrastructure
- Docker
- Terraform
- Kubernetes
- Jenkins
- Ansible
Concepts
- AI Agent Development
- Software Development
- API Development
- DevOps
Certifications
Courses and credentials I've completed.

CNCF - Certified Kubernetes Application Developer
2 September 2026
Cloud Native Computing Foundation's Certified Kubernetes Application Developer.
Credential ID: LF-5ridapwt2z

Anthropic - Introduction to agent skills
6 June 2026
Anthropic's Introduction to AI Agents Course

Meta - Introduction to Backend Development
Meta's Backend Development Course
Contact
Happy to talk about research, collaborations, or interesting problems.



