Applied AI · Singapore
CarlosEganaAppliedDataScientist·AIEngineer
I build agentic AI systems that plan, call tools and run multi-step workflows, covering retrieval, orchestration, evaluation and deployment.
- deploysDocker · AWS
- focusAgentic AI · MCP · RAG
- languagePython-first
About
Applied AI, from idea to production
I design, build and run LLM systems: agents, retrieval pipelines and the evaluation around them.
// carlos@ai-engineer:~$ cat profile.jsonc{"role": "Applied Data Scientist · AI Engineer","domain": "LLM systems: agents · retrieval · evaluation","focus": ["Agentic systems","MCP","RAG","LLM orchestration"],"language": "Python-first","background": "Information Systems & Control Management Engineering","status": "Shipping LLM systems to production"}Agents, not chatbots.
I build agents that plan, call tools and complete multi-step tasks, using function calling, MCP servers and multi-agent orchestration.
End-to-end ownership.
I take LLM systems from prototype to production: Python services, LLM integration, containerized deployment on AWS, monitoring, evals and usage analytics.
Built for real users.
I design around security and compliance requirements, cost control, and people who are not technical.
Skills
What I work with
The capabilities I use day to day, grouped with the tools behind them.
LLM Systems
- RAG pipelines & vector search
- Semantic chunking & reranking
- Prompt engineering & evals
- Structured outputs & context mgmt
- Claude API
- OpenAI API
- LangChain
- pgvector
- RAGAS
Agentic AI
- Multi-agent orchestration
- Tool use / function calling
- MCP — Model Context Protocol
- Agent evaluation & guardrails
- MCP
- Python
- FastAPI
Engineering
- Python services · FastAPI
- Docker & containerization
- AWS · ECR · App Runner
- Monitoring & usage analytics
- Python
- FastAPI
- Docker
- AWS
- Postgres
- Git
- PyTorch
- TensorFlow
- scikit-learn
- SQL
- Power BI
- Tableau
- Alteryx
- R
Approach
How I build agents
The pattern behind the agentic systems I build: retrieve, plan, act, evaluate.
01
Retrieve
RAG over the vector store grounds the model in real data
02
Plan
The LLM reasons about the task before it acts
03
Act
Tools resolved and executed through MCP servers
04
Evaluate
Evals and guardrails run before anything ships
→
Feedback loop
Each tool result updates the context
→
Deliver
Monitored and maintained in production
01
Retrieve
RAG over the vector store grounds the model in real data
02
Plan
The LLM reasons about the task before it acts
03
Act
Tools resolved and executed through MCP servers
04
Evaluate
Evals and guardrails run before anything ships
→
Feedback loop
Each tool result updates the context
→
Deliver
Monitored and maintained in production
Architecture pattern
# The pattern behind every agentic system I shipasync def run_agent(task: Task) -> Result:context = retrieve(task, vector_store) # RAG: ground the model in real dataplan = llm.plan(task, context) # reasoning before actingfor step in plan:tool = registry.resolve(step) # tool use via MCP serversresult = await tool.execute(step.args) # typed, validated I/Ocontext.update(result) # feedback loopreturn evaluate(context.output) # evals + guardrails before deliveryCredentials
Certified in cloud, AI and security
Credentials from AWS, Oracle, ISC², IBM, Google and others.
AWS AI Business Strategist
2026Amazon Web Services
AWS Certified Cloud Practitioner
2026Amazon Web Services
AWS Certified AI Practitioner
2026Amazon Web Services
OCI Generative AI Professional
2025Oracle
OCI Data Science Professional
2025Oracle
Certified in Cybersecurity (CC)
2025ISC²
Digital Transformation Management
2024Arizona State University
Data Science Practitioner · AI Practitioner
2024IBM
Alteryx Designer Core
2024Alteryx
Google Data Analytics Professional
2023Work
Selected work
Representative systems from my work in applied AI.
Flagship · Self-hosted AI agentAI that works for your organization, without your data ever leaving it.
Elkan is a self-hosted AI agent platform built for organizations that cannot send their data to a cloud chatbot. It runs entirely on the client's own infrastructure with local models, turns weeks of manual ISO 27001 and SOC 2 work into days of evidence-based auditing, and enforces governance by design: role hierarchy, full traceability and mandatory human review before any report leaves the system. Results are computed by code and auditable, not improvised by the model. Designed and developed end to end.
Elkan means protector or guardian in Mapudungun.
- built-in skills
- 84built-in skills
- compliance frameworks
- 8compliance frameworks
- ISO 27001:2022 controls
- 118ISO 27001:2022 controls
- governance roles
- 3governance roles
- Agentic AI
- Ollama
- MCP
- Python
- Self-hosted
Evidence-based auditing
A cyber audit agent assesses each control and delivers the Audit Report, Risk Assessment and SoA in Word and Excel.
Security operations
Ingests Nessus, OpenVAS, Trivy, Grype and SARIF scans and prioritizes findings with CISA KEV and EPSS.
Specialized agents
Each team composes agents with its own instructions, tools, skills and model, working on its own files.
Privacy and governance by design
Local models via Ollama, an isolated workspace per account, role hierarchy and mandatory human review.
Frameworks covered
- ISO 27001
- GDPR
- NIST CSF 2.0
- SOC 2
- PCI DSS
- NIS2
- DORA
- CIS v8
RAG over technical documentation
Retrieval pipeline over technical documentation, with semantic chunking, cross-encoder reranking and automated evaluation using RAGAS.
- RAG
- LangChain
- pgvector
Agentic workflow orchestration
Multi-step agents that use MCP tool servers, function calling and guardrails, deployed as containerized Python services on AWS.
- MCP
- FastAPI
- AWS
LLM evaluation pipeline
Automated evaluation of faithfulness, latency and cost, plus usage analytics and monitoring that non-technical users can read.
- Evals
- Docker
- Monitoring