Agentic Long-Runner
ReAct agent with 4 memory modes. Chunked vector retrieval fully recovers needle-in-haystack performance on 5,000+ word documents.
ML systems, data infrastructure, and applied AI. 19 open-source repositories with real benchmarks and production code. GitHub has everything.
Agentic Long-Runner
ReAct agent with 4 memory modes. Chunked vector retrieval fully recovers needle-in-haystack performance on 5,000+ word documents.
RenAIssance — GSoC 2025
Document AI pipeline for Renaissance manuscripts. Layout detection with LayoutLMv3, OCR, and post-correction on real historical pages.
MinIO + LakeFS Infrastructure
3-node distributed cluster with erasure coding, LakeFS versioning, and Prometheus + Grafana monitoring for ML datasets.
ArtExtract
CNN-RNN architecture for artwork classification. Trained on real art samples with confusion matrix validation and training curves.
CFN Biomedical Eval
Evaluation framework for biomedical NLP. Token-level and span-level benchmarks for clinical concept extraction.
NLP system for crisis-signal classification using social text, sentiment analysis, and geospatial mapping.
Real architectures from real projects. Each links to the source repo.
ReAct agent with 4 memory modes. Needle-in-haystack evals: 0% to 100% retrieval recovery.
Document AI for historical manuscripts. Layout detection, OCR, post-correction.
Distributed object storage with dataset versioning for reproducible ML experiments.
NLP crisis-signal classification with sentiment analysis and geospatial features.
Financial services for smallholder farmers. Credit scoring and mobile-first design.
A deep dive into the four critical bottlenecks slowing the robotics revolution and why general-purpose robots remain inevitable
A deep dive into how Partial Information Decomposition (PID) reveals how different modalities interact in AI systems, from redundancy to synergy