Open source AI projects

Open source AI projects and contributions by Wayner Barrios: inference engines, research benchmarks, model training tools, and geospatial platforms. Read the research publications behind the models and benchmarks.

Choose a starting point

If you want to run language or multimodal models locally on a Mac, start with the vLLM-MLX inference guide. It explains the serving workflow, the role of continuous batching and caching, and the measurements that matter when comparing models on your own hardware. The repository remains the reference for installation and supported features.

If you are evaluating the steps in a model's explanation, explore CRYSTAL. Its benchmark and metrics separate final-answer correctness from the coverage and ordering of reasoning steps. For work on instruction-guided visual features, the MoDA guide introduces the adapter and points to the model-family-specific training and evaluation resources.

The remaining projects cover training workflows, video research, development tools, and contributions to geospatial software. Entries marked as contributions describe my work within a larger project. Their repository links lead to the upstream projects, where you can inspect the code, documentation, and development history. The research publications provide context for the implementations associated with papers.

Open source

Built inthe open.

My research and tools first, then contributions to open-source communities — video, geospatial, and LLM infrastructure.

CRYSTAL

ECCV 2026 benchmark, dataset, and pip-installable metrics for evaluating ordered reasoning traces and training with process rewards.

CRYSTAL guide CRYSTAL code

6,372instancespippackage

MoDA

Official implementation of MoDA (ICML 2026) — instruction-guided channel modulation for fine-grained visual grounding in instructional MLLMs, with under 1% additional FLOPs.

MoDA guide MoDA code

ICML2026<1%extra FLOPs

vLLM-MLX

High-throughput, OpenAI- and Anthropic-compatible LLM and MLLM inference for Apple Silicon, with continuous batching, tool calling, vision, audio, and video support.

vLLM-MLX guide vLLM-MLX code

1.5k+stars400+tok/s

04DGX Spark Fine-tune LLMAn experimental end-to-end pipeline for LoRA fine-tuning, NVFP4/MXFP8 quantization, TensorRT-LLM export, and OpenAI-compatible serving on Blackwell GB10.4/8bit trainingGB10Blackwell 05Localizing MomentsOfficial PyTorch implementation of "Localizing Moments in Long Video Via Multimodal Guidance" (ICCV 2023) — identifies describable windows and prunes irrelevant segments before matching a query.ICCV202323stars 06OpenCode Power PackProduction workflows for code review, security, feature development, frontend design, and project memory, ported from Claude Code to OpenCode. Website: skills.waybarrios.com11skills450+stars 07ActivityNet ContributionBuilt and maintained the official ActivityNet challenge website — evaluation server, leaderboard, and GPU awards — for the leading video understanding benchmark, where I also filtered videos and measured the QA of the annotation process.114commits2016challenge site 08WorldMap ContributionHarvard CGA's WorldMap. I maintained the GeoNode 2.6-based platform: Django migrations, Dataverse/DataTables integration, and Dockerized deployment — 66 commits and 6 merged PRs.66commits6PRs merged 09GeoNode ContributionGeoNode, the leading open-source geospatial platform. I contributed Docker/GeoServer OAuth2 integration and the notification refactor (3 merged PRs), plus the Dockerized project skeleton (33 commits) and fixes across the GeoNode Docker stack.1.7kstars3merged PRs 10Hypermap Registry ContributionElasticSearch API, aggregations, and time faceting for Harvard CGA's Hypermap Registry — remote map services made easy for spatial data infrastructures.15PRs9merged 11MapProxy ContributionAdded dimension-layer caching for WMS and WMTS to MapProxy — the tile cache and WMS proxy at the core of OpenStreetMap infrastructure.666stars10commits 12GPT4All ContributionOne merged commit (2023): a DatasetDict dataset-handling fix for GPT4All's local LLM runtime.1commit2023merged

Building Wiqonn and researching multimodal perception, video understanding, and efficient inference.

Get in touch