nshkr-crucible

AI Firewall and guardrails for LLM-based Elixir applications

Explainable AI (XAI) tools for the Crucible framework

Deterministic tensor patch plans, patch application, and tensor path traversal for neural network model surgery.

Request hedging for tail latency reduction in distributed systems

Fairness and bias detection library for Elixir AI/ML systems

Experimental research framework for running AI benchmarks at scale

Interactive Phoenix LiveView demonstrations of the Crucible Framework - showcasing ensemble voting, request hedging, statistical analysis, and more with mock LLMs

Bumblebee, Axon, and Nx adapter layer for compiling Crucible tap plans into model runs, hooks, traces, and decode steering.

Model evaluation harness for standardized benchmarking—comprehensive metrics (F1, BLEU, ROUGE, METEOR, BERTScore, pass@k), statistical analysis (confidence intervals, effect size, bootstrap CI, ANOVA), multi-model comparison, and report generation. Research-grade evaluation for LLM and ML experiments.

Statistical testing and analysis framework for AI research

Metrics aggregation and alerting for ML experiments—multi-backend export (Prometheus, InfluxDB, Datadog, OpenTelemetry), advanced aggregations (percentiles, histograms, moving averages), threshold-based alerting with anomaly detection (z-score, IQR), and time-series storage. Research-grade observability for the NSAI ecosystem.

Canonical Elixir signal ontology for transformer forward-pass artifacts, tensor summaries, capabilities, and internal-control surfaces.

Intermediate Representation for the Crucible ML reliability ecosystem

Phoenix LiveView dashboard for the Crucible ML reliability stack

Data validation and quality library for ML pipelines in Elixir

Industrial ML training orchestration - backend-agnostic workflow engine for supervised, reinforcement, and preference learning. Provides composable workflows, declarative stage DSL, comprehensive telemetry, and port/adapter patterns for any ML backend. The missing orchestration layer that makes ML cookbooks trivially thin.

Nx SVD/SVF factorization primitives for neural network model surgery and TRINITY artifact export.

ML model registry for the Crucible ecosystem. Artifact storage, model versioning, lineage tracking, metadata management, model comparison, reproducibility, and integration with training pipelines for Elixir-based ML workflows.

Routing, gating, fusion, uncertainty, verifier, shared memory, and steering decision contracts over Crucible signal traces.