// open source · 5 repos · MIT Built in the open. Five reference projects for GenAI and agentic AI, built the way an enterprise has to run them: grounded, governed, measured and cheap to operate. Every number comes from a real run in the repo.
~/projects/README.md
Five repos, real results. // hover or tap a demo to play · recorded from each CLI's real output
genai · rag 01
▶
Enterprise RAG Reference problem Most RAG demos answer anything, cite nothing and ignore who is asking. 1.00 hit@4 retrieval
0 RBAC leaks · 17-case eval
1.00 correct refusals
Python Hybrid BM25 + vector RRF FastAPI Chroma RBAC
git clone enterprise-rag-reference ↗ agentic ai · langgraph 02
▶
Presales Copilot Agents problem Responding to an RFP takes days of reading, mapping, sizing and risk analysis. 12/12 requirements traced
296.2 person-days estimated
< 1 s per RFP
LangGraph Multi-agent Python Ollama / OpenAI / Anthropic
git clone presales-copilot-agents ↗ agentic ai · data 03
▶
Lakehouse Text-to-SQL Agent problem Letting an LLM write SQL against a warehouse risks wrong numbers and unsafe statements. 0.79 execution accuracy
8/8 attacks blocked
~9 ms per question
DuckDB sqlglot Semantic layer Snowflake Databricks
git clone lakehouse-text2sql-agent ↗ mcp · governance 04
▶
Enterprise MCP Toolkit problem MCP makes tools trivial; making them safe, scoped and auditable is the hard part. 3 governed MCP servers
15 security tests passing
$0 run cost
Model Context Protocol DuckDB Pydantic Audit log Claude / Cursor
git clone enterprise-mcp-toolkit ↗ llmops · evaluation 05
▶
Agent Eval Harness problem Teams ship prompt and model changes to agents based on vibes. 4 regressions caught
29% cheaper, still blocked
exit 1 CI gate
YAML suites LLM-as-judge SQLite Streamlit CI
git clone agent-eval-harness ↗