I'm a Data Scientist working on GenAI, LLM & agent evaluation, NLP, time series and GNN. Most of what I build in the open is about measuring AI systems rather than trusting them: how reliably agents call tools, which of a model's decisions are safe to automate, and what AI assistants actually tell people.

• 📦 Author of open-source packages on PyPI and npm with over 65,000 downloads, including APDTFlow for time series forecasting and Toolscore for LLM tool-calling evaluation
• 🧠 Released SafeDecide, a 140M decision model on Hugging Face with a certified error rate on the decisions it automates
• 🔌 Contributor to the Model Context Protocol: found two bugs in the reference memory server that silently lost saved data, and both fixes have shipped
• 🎓 M.Sc. in Statistics and Data Science, with a thesis on optimizing dynamic transportation networks using genetic algorithms
• ✍️ I write about this work on Medium

I'm open to collaborations and opportunities. Feel free to reach out!

• Languages:

  

• Frameworks & Tools:

  

• SafeDecide: A Decision Model That Knows When Not to Decide: A 140M model and Python package for typed decisions (yes/no, pick one, rate 0 to N) in the style of Jev. Instead of picking a confidence cut-off by eye, it certifies a threshold on labelled data: with probability at least 1 − δ, at most α of the decisions it automates are wrong, and the rest go to a person. The threshold comes from Learn-then-Test with exact Clopper–Pearson bounds. It serves the same /v1/systemone API as Jev, runs in the browser through ONNX and transformers.js, and calibrate_selective adds the same guarantee on top of any decision model you already use. safedecide calibrate re-certifies the thresholds on your own labelled data. Demo · Model · Method.
• Toolscore: LLM Tool-Calling Evaluation Framework: A deterministic, zero-API-cost Python testing framework that evaluates how well Large Language Models perform when making tool calls. It provides composite scoring across invocation accuracy, argument matching, sequence accuracy, and redundancy detection, with auto-detection support for OpenAI, Anthropic, Google Gemini, LangChain, and MCP. With over 12,000 downloads, it includes a full CLI (eval, compare, regression), GitHub Action integration, interactive debugging, multi-format reporting (HTML/JSON/CSV/Markdown), cost tracking, and a Pytest plugin for CI/CD pipelines. Articles on Medium: Toolscore for reliable tool-using AI · My agent said "saved". Nothing was saved.
• APDTFlow: A Modular Forecasting Framework for Time Series Data: APDTFlow is a modern and extensible forecasting framework for time series data that leverages advanced techniques including neural ordinary differential equations (Neural ODEs), transformer-based components, and probabilistic modeling. Its modular design allows researchers and practitioners to experiment with multiple forecasting models and easily extend the framework for new methods. With over 45,000 downloads. Read the article on Medium.
• FlowPrompt: Prompts as Code, Measured: A Python framework that turns prompts into typed classes with Pydantic-validated outputs — and then proves which version is actually better. compare() runs an A/B test across prompt variants and returns accuracy, latency, p-values and effect sizes; full experiments add production traffic splitting, sticky user assignment and multi-armed bandits. Provider-agnostic, with a pytest plugin for prompt regression tests in CI, OpenTelemetry tracing and multimodal support. With over 3,500 downloads.
• PromptBeacon: AI Brand-Visibility Measurement Engine: An open-source Python engine that measures whether AI assistants recommend your brand — across six providers in a single run, on web-grounded answers with real citations rather than model memory alone. It reports Share of Voice against your competitors, a stability score with confidence intervals (a single LLM answer is not a measurement), and glass-box funnel visibility that shows exactly where you drop out. Installs and runs keyless (promptbeacon demo "Nike"), and ships a pytest plugin and GitHub Action so a build can fail when AI stops recommending you. With over 4,500 downloads.
• Saylent: Open-Source AI Visibility Audit: Checks what ChatGPT, Claude, Gemini and Perplexity tell buyers about a brand. npx saylent audit example.com asks each engine the questions buyers ask, reads the pages they cite, tests the site the way AI crawlers see it, and writes an HTML report with prioritized fixes. Brand mentions are resolved in code rather than by a model, and every question is sampled more than once, so scores come as a range. Available as a Node CLI, a self-hosted Next.js + Supabase app, a GitHub Action and an MCP server. Apache-2.0. Docs · Sample report · Live demo.
• Project Management System with RAG: An AI-driven project management system using Retrieval-Augmented Generation (RAG) for enhanced efficiency in task execution, team collaboration, and decision-making. Read the article on Medium.
• VisualInsight: An app leveraging Google Generative AI (Gemini) to analyze images. This Streamlit-based web application allows users to upload images for analysis, storing both the original images and results in Amazon S3. Read the article on Medium.
• Bot Chat DeepSeek: An end-to-end pipeline of fine-tuning a Large Language Model (LLM) on AWS. This repository demonstrates how to prepare a custom dataset, fine-tune a language model, deploy it on AWS SageMaker, and interact with it via a Flask-based API.
• Football Scout RAG: An AI-powered football scouting agent designed to scrape data from Transfermarkt, providing advanced player statistics and comparisons.
• LLM-Code-Review 🤖: An automated code review tool powered by Large Language Models (LLMs), integrated as a GitHub Action. It provides real-time code analysis on every pull request, offering detailed feedback on security risks, performance bottlenecks, and code quality—right in your pull requests. No more waiting for reviews or missing critical issues.