Publicly available code samples

Non-proprietary projects completed outside of my employment.

0. Multimodal Applications

• Multimodal RAG - SigLIP for blazingly fast text and image encodings in the same embedding space + Qdrant as vector store + Gemini Flash to answer enriched queries.
• Intelligent Image Safety Triage - a cheap, specialist safety classifier that runs on 90% of Tier 2 images (shieldgemma-2-4b-it) and a large generalist model (gemini-2.5-flash) for a small slice of more complex Tier 1 flags - edge cases.

1. Agents

• Implementing different types of agents using multiple agentic frameworks

2. RAG

• Multiple RAG implementations using LangChain and LangGraph frameworks
• RAG Using Gradio App in HuggingFace Spaces
  • Technologies: Gradio, LangChain, ensemble retriever FAISS + BM25, re-ranking.
  • GitHub Repository: rag-demo-with-gradio
• RAG Using Multi-Page Streamlit App
  • Technologies: multi-page Streamlit app, LangChain, ensemble retriever FAISS + BM25, re-ranking, chat history.
  • GitHub Repository: RAG Demo with Streamlit

3. LLM Fine-Tuning (including from scratch)

• Fine-tuning every modality of Gemma 4
• Fine-tuning Llama 3.1 or a similar model like Ministral 8B
• Multi-GPU Fine-Tuning and Inference
• Preference optimization
• Continued pretraining
• RLHF
• Knowledge distillation
• Instruction Fine-Tuning of Llama 13B Model with DeepSpeed on a single or multiple GPUs.

4. Geoscience Application

• Predicting Direct Hydrocarbon Indicators (DHIs) using U-Net trained on stacked seismic attributes.
• Oil & Gas Well Toolkit - tools using GeoPandas to ingest and map surface well locations, Welly to plot well logs, PyVista to render 3D trajectories of horizontal wells, and Plotly to build an interactive HTML dashboard with rotatable 3D wellbore paths.

5. Recent Projects

Other non-proprietary projects. Note that being non-proprietary, this code doesn't reflect the complex and multi-faceted nature of projects that I normally complete at work. Content:

• NLP
  • Participation in two ACL (Association for Computational Linguistics) WASSA 2023 Shared Tasks
  • Prompt Engineering
  • ChatGPT, Gemini and other APIs
  • Transformers
  • Semantic search
  • Deep Learning (LSTMs and other DL models)
  • BoW models
• Other Machine Learning Topics
  • General machine learning algorithms and helper functions
• Certificates
  • Certificates for NLP and DL nano-degrees and other courses
• Visualization
  • I have used multiple visualization tools: Matplotlib, Plotly, Ggplot2, R Shiny, Tableau, Spotfire, etc.
  • D3.js visualization: website at https://agnedil.github.io/; code at https://github.com/agnedil/agnedil.github.io

6. Monte Carlo Simulation - Solution for the Monty Hall Problem

Monte Carlo Simulation for the Monty Hall Problem - empirical experimental confirmation of the famous counter-intuitive answer to the three-door problem: always switch your answer.

7. Evaluating Quality of Machine Translation

Machine Translation Quality Evaluation using reference-based and reference-free metrics, as well as LLM-as-a-judge approach.

8. Detecting Fake Customer Reviews of Products

Detecting fake customer reviews of products on Amazon. This is a machine-learning based approach that allows to determine fake reviews automatically. Use case: help buyers understand whether the reviews for a specific product are authentic or fake to make informed decisions about what products to purchase. The methodolgy also determines the percentage of fake reviews.

9. Code Generation

Code for my doctoral dissertation - Boosting the Code Generation Capabilities of Small Language Models (SLMs) Using Agents. Includes:

• a framework to automatically run and evaluate multiple small language models on 4 code generation evaluation datasets,
• hyperparameter tuning,
• post-processing of generated code,
• SLM fine-tuning,
• experiments with consecutive improvements.

10. Python Developer Coding Interview Preparation

Practical Guide to Prepare for a Python Coding Interview collected from multiple sources. Includes:

• Coding challenges from most of the FAANG companies.
• Algorithms and data structures as examples of Python code.
• Big O notation explained based on practical examples of Python code.
• Other coding resources

11. Other Old Project

Non-proprietary projects completed long time ago. Somewhat outdated and simplistic solutions - keeping for reference. Content:

• NLP
  • Embeddings
  • Topic modeling
  • Conversational AI
  • Other miscellaneous projects
• Machine Learning
  • General machine learning algorithms and helper functions
• Computer Vision
  • Semantic segmentation, image denoising, classification and other image analysis methods
• Cloud Technologies
  • Hadoop, MapReduce, MLlib, Giraph, Apache Spark, etc. x