Arad Fadaei

Ethos is a Python runtime I’m developing to provide a shared foundation for AI agents. It brings model integration, persistent sessions, tools, and configurable personas together around workspace-scoped context. The runtime connects to OpenAI, Google, and Ollama models, with a command-line interface and an authenticated REST API built on FastAPI.

The architecture separates agent execution and application state from the interfaces that use them. Both the CLI and API operate through the same service layer, while capabilities contribute tools and instructions through a common interface. This keeps behaviour consistent and allows the runtime to grow without coupling its core logic to individual clients or tools.

A central focus is controlled, recoverable execution. Agents can inspect project files, request approval for changes, and run approved commands within a native sandbox. Durable approval records and explicit recovery for interrupted actions help prevent accidental replay, making permissions and failure handling part of the runtime itself.

Ethos also includes an evaluation harness that exercises models against task-completion and security scenarios using the real agent runtime. Reports track pass rates, tool usage, latency, and token consumption to support evidence-based model comparisons. The project is currently in alpha, with ongoing work on the core runtime and a longer-term direction towards workflows and orchestration.

aethellib is a Rust library for composable text generation from target-specific TOML corpora, with built-in source provenance tracking. The library focuses on loading and validating attributed documents into a single-target corpus, then pooling values by section and field so generation rules can operate predictably and transparently.

The generation engine is built around composable rules rather than a rigid DSL. You can combine primitives like literals, picks, and concatenation to construct outputs step by step, while the engine stores prior results in context for reuse across rules. This makes complex generation flows easier to reason about and extend.

A core differentiator is provenance as a first-class concept. Generated values preserve traceable source metadata, so outputs can be inspected and explained back to the originating corpus values. That traceability is especially useful when generation quality, data lineage, or debugging clarity matters.

The project is currently at version 0.8.4 and under active development. It is published on crates.io and can be installed with cargo add aethellib. For a complete reference workflow, the repository includes a runnable example that demonstrate corpus loading, rule composition, and context-based output retrieval in practice.

Pubsie is an open-source Node.js library I developed to simplify the process of parsing EPUB files. It allows developers to efficiently extract essential components such as metadata, chapters, and other resources from EPUB files within JavaScript applications.

The library features an event-driven architecture, providing robust error handling and facilitating asynchronous operations during the parsing process. To enhance performance, Pubsie includes a caching mechanism that stores parsed data, reducing the need for repeated parsing and improving data retrieval speed.

In building Pubsie, I focused on compliance with EPUB specifications to ensure broad compatibility with standard-compliant files. The project is available on npm, complete with detailed documentation and usage examples to assist developers in integrating it into their applications.

Capstone Project | Team Lead & Scrum Master

Moneymap is a full-stack personal finance tracking application developed as part of a capstone project. As the team lead and Scrum Master, I directed a team of developers and designers through the full Agile development process, from planning and architecture to final delivery.

The core of Moneymap is an AI-powered assistant designed to help users make sense of their finances. Users can interact with the assistant to ask questions, get personalized insights, and receive context-aware summaries based on their spending habits, budgets, and trends. This conversational layer made financial data more accessible and actionable.

In addition to leading sprints and team coordination, I contributed to both front-end and back-end development. The application features automated expense categorization, budget tracking, and data visualizations to give users a clear picture of their financial health. We conducted multiple rounds of usability testing to refine the interface and improve the user experience.

Moneymap demonstrated our team's ability to integrate AI with financial tools to create a modern, responsive application that turns complex data into clear, useful guidance.

NLP-PhishGuard is a supervised natural language processing project focused on phishing email detection. The core challenge addressed in this work is that many modern phishing messages are professionally written and semantically similar to legitimate business communication, making rule-based filtering alone unreliable in real-world settings.

The pipeline was designed end-to-end to be reproducible and practical, covering ingestion of labeled corpora, preprocessing, stratified train-validation-test splitting, transformer fine-tuning, checkpointing, and final evaluation. This structure made it possible to measure generalization on held-out data while maintaining a clear engineering workflow for retraining and iteration.

Beyond predictive performance, the project emphasizes interpretability and usability. Token-level attribution helps inspect why a message was flagged, and the final model is packaged for both command-line usage and lightweight web-based inference, making it easier to integrate into practical security workflows.

CV_SPAC is an end-to-end computer vision project for automated parking access control using license plate recognition. The system is motivated by the limitations of manual checkpoints, keypad-based entry, and card-based access methods, which add recurring operational overhead and can still fail in routine use because of process friction or human error.

The implementation combines a lightweight plate detector, OCR with preprocessing, text normalization, and resident-database matching with both exact and fuzzy strategies. This integration focus is central to the project, since robust access-control behavior depends on pipeline quality across stages rather than any single model metric.

The final output produces grant-or-deny decisions with confidence scores and supports reproducible evaluation against labeled ground truth. The project is packaged as a configurable CLI pipeline so it can be tested, tuned, and deployed with predictable behavior in real access-control scenarios.