Mihai Chirculescu

An application framework that simplified developers' work and automated the work of other departments exchanging documents for validation with the company's business customers via email for their specific use case.

From a technical perspective, the web app generated Ext JS controls and UI views using the database tables. The purpose was to simplify developers' work while ensuring security and different user role permissions. It was highly configurable and communicated with the back-end via RESTier.

A collateral management and optimization tool built with a microservice architecture using IBM's CPLEX. I have implemented new features and optimized the multi-million collateral allocation tool.

I solely built an NLP (AI) semantic search component that runs locally, without any external calls, and returns the result in less than 0.2 seconds. The component uses an open source ML model that generates embeddings on which normalization is performed and cosine similarity is done to establish the order of the items/options. The tool runs on Windows using ONNX for inference and is completely integrated into our desktop app.

I developed a Python module to steer large language model (LLM) responses toward a certain topic or subject and enhance capabilities (e.g., making it provide correct responses to tricky logical puzzles more often). This model is a practical tool for using activation engineering by adding steer vectors to different layers of an LLM. It should be used along with the Transformers library.

I fine-tuned multiple open-source models based on Mistral, Llama2, and Yi for reasoning tasks and long system prompts that can be used in a corporate environment. I published them on HuggingFace under open-source license.

I made a React component that extends and performs semantic similarity search using a small, quantized machine learning (ML) model that runs on the client side. I distilled and fine-tuned multiple small embedding models (based on my own synthetic dataset) to work with this component.

I made a live multiplayer trivia game where users can bid for the subject of the next question. A local LLM will generate this next question. It was built using Answer AI's fastHTML while in beta.

I made a tool to help me better understand LLM samplers and create LLM samplers that can rewind and revise generated tokens. Backtrack Sampler is a framework for experimenting with custom sampling algorithms (strategies) that can backtrack/undo/rewind/reverse the latest generated tokens.

I made a Twitter bot that provides YouTube video chapters by calling an external LLM. It integrates with Google's Gemini under the hood, and all the user needs to do is tag @TimeStampBuddy with a link to the YouTube video. The bot will reply within a few minutes.