Abel Dantas · AI Agents & Automation

I build governed AI-agent systems: MCP tools, RAG over knowledge graphs, evaluation loops, and human-in-the-loop approval. Agents in production, daily. PhD in distributed systems.

Claude (SDK + Code)MCPLangGraphn8nWindmillRAGNeo4jeval pipelineshuman-in-the-loop

Havenware — Governed Agent Fleet

Claude · MCP · per-agent memory · knowledge graph · governance layer

havenware.devApproval queueTapering reviewThe fleetTraining dataMission control
havenware.dev Your own governed AI fleet — a team of AI agents that runs your operations, and nothing they do ships without your yes.
Approval queue Every agent action is a proposal — quick yes/no calls and deep reviews, tiered by estimated review time. Nothing executes without an explicit human verdict.
Tapering review Human verdicts become ground-truth labels; settled cases harden into deterministic rules, so the share of actions needing review falls week over week.
The fleet Five role agents — inbox, outreach, social, finance, research — with live sessions, current objectives and what is waiting on the owner.
Training data Approvals and corrections land on the immutable trail and become export-ready, per-agent fine-tuning datasets. Self-hosted; no provider training on user data.
Mission control LLM spend against budget, service health and the decisions that need a human — the whole fleet on one screen.
A product built from my own agent operating system: a fleet of AI agents that runs an organization’s inbox, outreach, social, finance and research, under a governance layer where nothing ships without a human yes. Running in production, daily. Review load down 41% in eight weeks — 214 decisions graded, 93.4% owner agreement.

havenware.dev

CIM Alto Minho — Government Decision-Support AI

RAG · knowledge graph · AI agents · spaCy NER · Python · FastAPI · Neo4j

Governed agentsActivity dashboardDocument intelligenceSemantic searchPublic channelFederation
Governed agents An internal RAG agent and a public WhatsApp assistant, each with its own on/off switch; answers cite sources with relevance scores.
Activity dashboard 12,872 documents, 8,912 graph entities and agent-query volume across a 10-municipality federation, with live source syncs.
Document intelligence spaCy NER highlights entities inline; extracted metadata with SHA-256 integrity and related documents resolved through the knowledge graph.
Semantic search Query interpretation, relevance-ranked results and municipality/type/year facets over the whole archive.
Public channel An anonymous WhatsApp assistant restricted to public information — no sign-up, every answer names its source document.
Federation Each municipality keeps its own graph node; an interoperability protocol answers cross-cutting queries without centralizing the data.
An AI decision-support platform for an inter-municipal government of 10 municipalities, live in production: the public document archive ingested into a knowledge graph, explorable visually, searchable semantically, and answerable through AI agents that always cite their sources. 84.7% of answers ship with cited sources; the public WhatsApp channel runs at 94.2% satisfaction.

CONFECOOP Federation — Governed AI Assistant

LLM · RAG · MCP · Neo4j · human-in-the-loop · eval dataset

Institutional assistantKnowledge baseHuman-in-the-loopUnder the hood
Institutional assistant Grounded answers with a traceability panel: cited statute articles, per-source relevance scores, and the query logged to the audit trail.
Knowledge base The confederation's documents as a Neo4j knowledge graph with vector retrieval feeding the assistant.
Human-in-the-loop LLM document classification behind a review queue — approvals and corrections become labelled training data.
Under the hood The live retrieval pipeline: Gemini embeddings, Neo4j vector search and scored passages, run against production.
The governed AI assistant for CONFECOOP, the national confederation for Portugal’s cooperatives and social-sector enterprises, serving its 85+ member organizations. A RAG assistant over the institution’s own archive that always cites its sources, with document classification kept honest by a human-review queue. A 4,213-record hash-chained audit log verifies end to end.

SIGO — Agents on a Ledger

MCP server · approval-gated ledger · least-privilege agent keys

Human approval gateAppend-only ledgerRead-only AI agentsCounterparty CRM
Human approval gate A pending €31,450 movement with its full transaction log and grant attribution; balance records show current vs. pending impact before anyone signs off.
Append-only ledger 38 movements across varied counterparties with approved, rejected and pending states — reversals are compensating entries, nothing edited in place.
Read-only AI agents Agents reach the ledger through an MCP server under scoped, auto-minted keys; no write tools exist.
Counterparty CRM Contacts with owners and follow-ups across the organization's counterparties.
A financial back-office platform, in production, where AI agents read the books through an MCP server and money only moves with a human sign-off, on an append-only ledger. Adopted across public and social-economy institutions.

sigo.so

EU AI Act Advisory Forum (European Commission)·Agents in production, daily·PhD in distributed systems·Published GNN research