Santander AI Open Source
Open-source AI org from Banco Santander AI Lab (Madrid). Publishes tools for small models, harness engineering, evolving agents, responsible AI, MLOps, and graph ML for financial services. All projects ship under permissive licenses (mostly Apache-2.0) and use synthetic or anonymised data only — no real customer data. Notable as a regulated bank running a formal Open Source Programme Office (OSPO).
Links
Description
Run by an OSPO with a transparent two-track release review:
- Fast Track — forks, generic tools, tutorials, datasets, SDKs without business logic; OSPO Lead + automated scans, SLA < 4h.
- Full Track — models, IP-bearing frameworks, code that touched internal data; FOSS Review Board (OSPO Lead + Legal + CISO + Architect), SLA 2–4 weeks.
Download or use
- GitHub org: https://github.com/SantanderAI
- Governance: GOVERNANCE.md · CLA on first PR.
- Contact: opensource@gruposantander.com
🧩 Featured projects
| Project | What it is |
|---|---|
ralph | Configurable Bash/PowerShell loop running an AI coding CLI with a fresh session each iteration — a packaged Ralph loop. |
ralph-vault-skill | Skill to generate the knowledge vault for projects using the Ralph loop. |
autoguardrails | Alignment-research scaffold (autoresearch-style) for LLM guardrails over a single policy.md surface. |
mech-gov-framework | Mechanical Governance for LLM decisions — model-agnostic regimes, hard gates, governance metrics for high-stakes decisions. |
auto-bayesian | Config-driven, interpretable Bayesian network training for relational tabular data. |
gen-fraud-graph | Synthetic fraud graph generator (scales to 100M+ accounts) for benchmarking graph fraud detection. |
genetic-algorithm | Dependency-free Python GA engine with pluggable fitness — search core for an LLM autoresearcher. |
linear-adapter-trainer | Train linear embedding adapters (triplet loss) to align retrieval embeddings with queries (RAG). |
llm_bridge | Tiny vendor-neutral LLM client — one interface, pluggable OpenAI/Bedrock/Gemini adapters. |
causal-perception-implementation | Causal perception research — competing structural causal models for fair credit decisions. |
mutatis-mutandis | Situation testing for discrimination analysis with counterfactual comparators. |
sota-stressed-datasets | Benchmark datasets republished in “stressed” form to evaluate ML/LLM robustness. |
Reasoning for
Direct overlap with this vault’s working themes. ralph is a production-grade implementation of the loop-engineering idea — fresh session per iteration is exactly the context-rot mitigation behind Loop Engineering and the Harness Engineering practice; ralph-vault-skill auto-generates a project knowledge vault, paralleling LLM Knowledge Bases / OpenKB. autoguardrails and mech-gov-framework are concrete patterns for the “keys not prompts” / hard-gate governance lesson (see Building an AI Second Brain). Worth mining for harness and agent-governance ideas from a regulated-bank perspective.
Alternatives considered
- Archon, Superpowers, gstack — other harness/agent-skill stacks; Santander’s angle is governance + responsible AI under banking constraints.
Resources
- Related: Harness Engineering, Loop Engineering, Autoresearch, OpenKB
Template: tool