Data Scientist | Full Stack Developer | Machine Learning Engineer
I am a Machine Learning Engineer with 6 years of experience building production AI systems across NLP, computer vision, and agentic applications.
Currently, I am a Consulting Applied AI Engineer at Everpure .
In my prior roles, I served as a Senior Data Scientist at Agency Enterprise, ML Team Lead at the Frontier Development Lab (in partnership with ESA & NVIDIA) and a Full Stack / Machine Learning Engineer at Dwight Funding. I originally trained in Physics and Astronomy (UMD), where I interned at NASA Goddard. After which I earned an M.S. in Data Science from NYU.
When I'm not coding, you'll find me running, geeking out on Star Trek, or tinkering with local AI.
Python • PyTorch • LLMs & Agents • RAG • Hugging Face • TypeScript • FastAPI • Docker
A proof of concept showing how activation steering can subliminally guide users towards products or brands. Demonstrates how AI responses could be influenced by advertisers.
Hackathon winning project. Aligned and steered models through interventions to model weights based on diffing and and continual pre-training.
Built reinforcement learning training infrastructure for vision-language models, enabling iterative improvement of object detection performance through policy optimization and custom IoU-based reward shaping.
Browser extension to monitor and reduce addictive scrolling on X.
A Foundation Model for the Solar Dynamics Observatory.
Fast local assistant hosted on a Raspberry Pi.
Implemented LLaVA, a vision language model, in MLX, a framework for building AI applications for Apple devices.
Found and resolved bugs in the LitGPT project and added a new feature.
Flood Disaster Reporting by Web-based Retrieval Augmented Generation with an LLM.
Semantic search for Hugging Face datasets.
A competition to detect field boundaries in satellite imagery.
Training a remote sensing model on a Jetson Nano 2GB.
And many more!
Extracting "brand vectors" for 24 companies from Qwen2.5-7B and adding them to the residual stream mid-generation. How the vectors are made, how well they work, and why I think this is how LLM ads will actually happen.
I tried to catch Qwen3-32B reward hacking by watching its activations with an off-the-shelf interpretability direction. It didn't work, and the way it didn't work is the interesting part.
Teaching a small open-source VLM to detect objects on an image using another image as a prompt.
Training an RL agent to play Slither.io.
I've been using Cursor's Tab model, but it doesn't work as well as I'd like for how I start my data science flows. Here's what I did to make my life easier.
You can email me directly at nkasmanoff@gmail.com or use the form below.