/projects/boundeddata
A production-oriented TypeScript framework for governed data agents: the model plans and explains, deterministic tools compute, and governance gates every bounded, read-only query — with sampling, verification and full audit traces.
Enterprise data agents are risky when models can write arbitrary SQL or cite unverified retrieval. Teams need bounded, read-only querying with deterministic verification and audit traces.
Bounded Data Agent separates model planning from deterministic tools: semantic query, hybrid search, sampling, profile SQL, bounded SQL, verification, evaluations, and audit trace are distinct stages.
The system must constrain SQL scope, preserve database safety, verify sampled answers, support vector retrieval, and explain how every answer was produced without exposing uncontrolled database access.
The architecture treats the model as planner and explainer while typed tools perform computation. Governance is built into the retrieval and SQL path rather than bolted on after generation.
This shows Stefan's command of governed RAG, Postgres/pgvector, Drizzle schemas, LangGraph-style flows, SQL safety, verification harnesses, and audit-focused data-agent design.
Bounded Data Agent is a governed analytics system for model-assisted querying. Its depth is in keeping the model as planner and explainer while deterministic tools do retrieval, SQL, verification, and audit.
The agent starts with a business semantic layer and hybrid retrieval before raw SQL. That keeps questions aligned with domain concepts instead of table spelunking.
Bounded SQL is scoped, read-only, and policy-approved. The model can request computation, but it does not get unrestricted warehouse authority.
Sampling, profiling, and verifier stages make answers harder to fake. The system checks data quality and result plausibility before the final response.
Questions, plans, tool calls, SQL, evidence, and answers are recorded so an enterprise reviewer can reconstruct how a conclusion was produced.
Keeps the planner, tool executor, policy objects, and answer contracts typed end to end.
Fits the HTTP and CLI runtime while staying close to the TypeScript agent framework.
Acts as the governed warehouse, catalog store, document index, and audit-friendly source of truth.
Keeps semantic retrieval inside Postgres so vector search and relational governance share one boundary.
Provides typed schema, migrations, and seeds without hiding the SQL policy layer.
Makes the data agent reproducible with Postgres and local services available for testing.
Gives fast unit and evaluation checks for policy, query, and verifier behavior.
Matches the explicit retrieve, sample, execute, verify, answer pipeline shape.
Asks the business semantic layer before touching raw warehouse tables.
Combines keyword and vector retrieval across documents, graph data, and embeddings.
Chooses representative data before the agent attempts broader analysis.
Inspects distributions and data quality so answers can state limits honestly.
Allows only guard-approved, scoped, read-only SQL to execute.
Checks results deterministically before the answer writer responds.
Scores fixture runs against thresholds so quality can be tracked.
Records the question, plan, tool calls, SQL, and evidence for review.