Researching and building practical, open-source AI systems.
๐ Open to remote opportunities in software engineering, AI systems, and developer tooling.
Building systems that make software and AI more readable, reliable, and verifiable.
| Project | What makes it different |
|---|---|
| Semauri. | Where human language meets programming. A natural-language-inspired, type-safe language bridging human creativity and AI-powered development, with a Rust compiler, source-aware diagnostics, and explicit control over side effects. |
| TrainLens | Turn training runs into next experiments. Deterministic notebook diagnostics, visualizations, and evidence-backed recommendations, with optional LLM and agent workflows. |
| ASTScribe | Understand ML notebooks without executing them. Static analysis traces model setup, training, evaluation, and cell dependencies back to their original source lines. |
| LastLight | Reliable retrieval when infrastructure fails. Standard-library-only, offline knowledge search with inspectable sources and evidence-based refusalโno cloud, embeddings, or vector database. |
A tiny, slightly absurd, surprisingly useful Python library: turn -- into a proper em dash (โ) in text, Markdown, and Word documents. Because the little things matter.
Try it!
pip install emdashify
from emdashify import convert convert("Hello -- world", compact=False) # 'Hello โ world'
| Project | Focus |
|---|---|
| Medicine Pack Extractor | Reproducible Python/Jupyter pipeline that extracts official medicine documents from Spain (AEMPS/CIMA) and Italy (AIFA), preserves provenance, and builds bilingual knowledge packs for LastLight. For retrieval researchโnot medical advice. |
Python Package Portfolio
Small, practical Python tools for AI evaluation, data quality, retrieval, multimodal analysis, visualization, and generative-audio workflows.
| Package | Research / engineering focus | Downloads |
|---|---|---|
| context-dedup | Context deduplication for reducing redundant content in LLM and agent workflows. | |
| evidenceflow | Deterministic analysis of how retrieved and tool-generated evidence propagates through AI-agent traces. | |
| rag-chunk-audit | RAG chunk-quality auditing for segmentation and retrieval analysis. |
| Package | Research / engineering focus | Downloads |
|---|---|---|
| embedding-drift-lite | Embedding-distribution drift detection and diagnostic inspection. | |
| parametricbench | Provider-independent regression benchmarking for LLM and VLM systems. | |
| promptshield-llm | Prompt-injection, unsafe-instruction, and input-risk screening for LLM pipelines. |
| Package | Research / engineering focus | Downloads |
|---|---|---|
| metaclean-vlm | Metadata normalization for image and vision-language model datasets. | |
| visual-patch-audit | Image-patch auditing and visual-signal inspection for VLM workflows. | |
| vlm-occlusion | Grid-based black-box occlusion analysis for VLM claim sensitivity. | |
| vlm-prior-probe | Counterfactual black-box evaluation of whether VLMs follow visual evidence or learned priors. |
| Package | Research / engineering focus | Downloads |
|---|---|---|
| matplotlib-dark | Automatic dark mode for Matplotlib with ready-to-use themes and safe temporary styling. |
| Package | Research / engineering focus | Downloads |
|---|---|---|
| egypttranslit | Egyptological transliteration to clean Unicode from Python or the command line, with MdC conversion, normalization, validation, and diagnostics. |
| Package | Research / engineering focus | Downloads |
|---|---|---|
| text-to-music-prompt-structurer | Structured prompt extraction for text-to-music generation workflows. |
| Guide | What's inside |
|---|---|
| Neural Audio Theory | A practical and technical guide to AI music generation, signal processing, embeddings, model architectures and training. Live guide. |
| edujbarrios-ai | AI algorithms from scratch in C, Python, Rust and R, with explanations, documented implementations and tests. |
| edujbarrios-ui | A compendium of copy-ready React and TypeScript frontend components for AI products, dashboards and developer tools. Component gallery. |
| Project | Research focus |
|---|---|
| Cablegram | Token-efficient communication for humans and AI agents, with auditable context compression and invariant checks. |
| NEXO | Versioned communication protocol for AI agents, with reversible text/binary encoding, capability negotiation, media transfer, and token-aware JSON fallback. |
| Suffice | Finding the minimum successful token budgets for AI agents under explicit quality constraints. |
| MAVERICK | Cognition-inspired, multi-agent visual reasoning with an interpretable perceiveโdescribeโcritiqueโrefine loop. |
| MOSS-VL MCP | Experimental, unofficial MCP server bringing MOSS-VL visual observations to coding agents for screenshots, UI comparisons and test recordings. Requires a separately configured inference backend for visual analysis. |
| music-to-text | Local-first audio analysis: acoustic feature extraction, structured metadata, A&R notes, PR pitches, playlist descriptions, and sync licensing copy. |
| Benchmark | Evaluation focus |
|---|---|
| LastLight Bench | Reproducible offline retrieval evaluation: ranking, answer/refusal calibration, bilingual behavior, routing, latency and memory usage. |
A couple of lightweight API clients worth keeping in the toolbox.
API Clients
| Library | Language | Focus |
|---|---|---|
| llm7R | R | Lightweight client for the LLM7.io API, with chat, streaming, model discovery, JSON mode, tool calling, data-frame analysis, vision, and image/video generation. |
| llm-ts-api-wrapper | TypeScript | Zero-runtime-dependency client for OpenAI-compatible APIs, covering chat and Responses APIs, streaming, embeddings, models, tool calling, retries, timeouts, and typed errors. Designed for direct source integration; not published on npm. |
Languages
AI Systems & Tooling
- I Created a Complete Educational Guide on AI Music Generation
- I Built My Own Documentation Site Builder - Here's Why and How
- I Built a Notebook Engine for C, and the Potential Is Incredible
OS & Development