Human-approved workflow agents
Turn a request into a structured proposal, pause for review, then allow only the exact approved action to reach a business tool.
FADEL SANTIAGO / FREELANCE DATA & AI ENGINEER
AI agents that fit the way your business works.
I help companies assess where AI is useful, design controlled workflows, integrate models with existing data and tools, and build the software needed to operate them reliably.
PRACTICAL AI ENGINEERING
Start with a business process, a decision or a knowledge bottleneck, not an agent for its own sake.
A useful first step may be an audit, a small proof of value, a complete implementation or focused engineering support. The scope should match the operational risk and the evidence available.
01 / FROM QUESTION TO OPERABLE SYSTEM
One engagement, six controlled stages.
Each stage produces something reviewable before more complexity is introduced.
- 01
Audit & opportunity mapping
Map the current process, systems, data, failure points and manual decisions. Separate genuine AI opportunities from simpler automation or product work.
- 02
Feasibility & strategy
Define the target outcome, constraints, data readiness, model choices, evaluation plan, operating boundaries and an implementation route.
- 03
Workflow & solution design
Design the human and system journey, tool permissions, approval gates, retrieval flow, interfaces, contracts and observable states.
- 04
Implementation & integration
Build the agent or AI workflow, connect approved data and business systems, and create the supporting APIs or application interface.
- 05
Validation & deployment
Test expected, failing and unsafe paths; document limitations; package the system for the agreed environment and handover.
- 06
Maintenance & enablement
Monitor the implemented scope, improve controls, support operations and transfer the knowledge required for your team to own it.
02 / USE CASES
Where controlled AI can remove real friction.
The best use case is narrow enough to evaluate and valuable enough to change how work gets done.
RAG & document intelligence
Search internal documents semantically, generate from bounded evidence and keep citations visible so answers can be checked.
Operational triage & routing
Classify incoming work, extract structured fields, identify missing information and route cases while keeping exception handling explicit.
Knowledge assistants inside existing tools
Connect a focused assistant to the interfaces, data sources and permissions people already use instead of creating an isolated chatbot.
Workflow and decision support
Combine rules, model outputs and human checkpoints to reduce repetitive hand-offs without hiding responsibility.
AI adoption audit
Assess processes, data, constraints and organizational readiness before committing to a platform or large implementation.
ENGINEERING CONSTRAINTS
Useful means controlled, inspectable and maintainable.
- Authority
- The model proposes only what it is allowed to propose. High-impact actions can require explicit human approval.
- Evidence
- RAG answers should expose validated sources and distinguish missing evidence from confident wording.
- Data boundaries
- Inputs, storage, providers and access paths are chosen from the real confidentiality and infrastructure constraints.
- Reliability
- Contracts, tests, failure states, retries and audit trails matter as much as the prompt.
- Operations
- Deployment, monitoring, maintenance and team ownership are planned as part of the system, not after the prototype.
03 / VERIFIED PORTFOLIO EVIDENCE
The engineering is visible.
These are working portfolio demonstrations with documented scope and limitations. They are not presented as client deployments.
HUMAN-APPROVED AI WORKFLOW
RELAINE
A local agentic workflow where an Ollama model creates a strict proposal, execution pauses for approval, and persistent tools record the result and audit history.
103 non-Ollama regression tests and 2 real Ollama tests passed for the documented MVP release.Examine RELAINE ↗RAG DOCUMENT INTELLIGENCE
ATLAST
A local-first document workflow for PDF and TXT ingestion, semantic retrieval, bounded Ollama generation and citations validated against retrieved passages.
40 automated tests passed; the published retrieval benchmark is explicitly small and synthetic.Examine ATLAST ↗04 / ENGAGEMENT
Choose the smallest useful starting point.
Every route uses the same project inquiry. Describe the business change, current systems and constraints; the technical shape can follow.
Initial audit or feasibility discussion
Clarify the process, opportunity, data readiness and risks before choosing an implementation.
Request an initial audit ↗Complete implementation project
Commission a defined AI integration or workflow from architecture through delivery and handover.
Discuss a complete project ↗Freelance engineering mission
Add focused AI, data or application engineering capacity to an existing team and delivery plan.
Discuss a freelance mission ↗WIDER PRACTICE
The full Data & AI engineering scope.
- Consulting, Audits & Strategy
- AI & Agentic AI Engineering
- Data Engineering, Big Data & Business Intelligence
- Technical Support, Maintenance & Training