Most "AI agent" demos stop at the prompt. My work starts where it gets hard: an agent with real credentials, on a real server, that must never break anything.
| Discipline | How it shows up in the code |
|---|---|
| Spec-driven | Every feature ships as spec → plan → tasks before a line of code |
| AgentOps | Command allowlists, approval gates, sandboxing, least-privilege SSH, audit ledger |
| LLMOps | Model judgment (alert severity, secret redaction) pinned by versioned golden cases in CI |
| Fail-closed | Missing config, ambiguous state or unknown command → refuse, never guess |
| Secret hygiene | Zero credentials in code, logs, prompts or reports — enforced by CI scanning |
I treat an AI system as a lifecycle, not a demo — and each stage has its own operational discipline:
| Practice | In one line |
|---|---|
| Agile delivery | Iterative increments, short feedback loops, working software over ceremony |
| Spec-driven | spec → plan → tasks, each task small, reviewable and independently testable |
| MLOps | Reproducible data pipelines, versioned corpora, lockfile-pinned environments |
| LLMOps | Golden-case eval suites in CI, tracing/observability, regression-proof prompts |
| AgentOps | Risk-classified actions, approval gates, sandboxing, full audit trail |
| Docs as contract | Runbooks, requirement docs and ADR-style decisions live beside the code |
|
Multi-source ETL, freshness & volume checks, load-control tables, scheduled orchestration and real-time incident alerting. |
RAG over vector stores, NLP & deep learning (Word2Vec, RNN/LSTM), computer vision (YOLOv8), ensembles. |
Linux servers, WSL2, systemd, circuit design and embedded controllers — hardware-level intuition applied to software. |
Rendered in this repo from public GitHub data — no third-party stats services, no token, no JavaScript.
Refreshed daily by a GitHub Action. Build it yourself:
python tools/build.py
The full arc of an AI system, built end to end: a public-source OSINT pipeline on Manuelita S.A. — a 160-year-old Latin American agro-industrial group operating in 3 countries and exporting to 65+ — turned into a semantic knowledge base, then into a RAG agent with memory, and finally productized on an Agentic Operating System that answers from Telegram and WhatsApp.
Stage What was built M1 · MLOps foundation Public-source ingestion, NLP cleaning with spaCy, semantic corpus + Q&A M2 · RAG agent LangChain + ChromaDB vector retrieval, conversational memory, Streamlit UI M3 · AgentOps Productized on OpenFang (Rust Agent OS) over WSL2, Ollama/Gemini models, live on Telegram + WhatsApp, LangSmith observability, t-SNE embedding analysis
Python 3.11·LangChain·ChromaDB·spaCy·Gemini / Ollama·OpenFang·LangSmith·StreamlitM.Sc. in AI & Data Science — Universidad Autónoma de Occidente. Built with a 4-person team.
| Project | What it does | Stack |
|---|---|---|
| os-system-agent | Controlled agent that monitors ETL jobs on remote servers so operators stop SSH-ing into production for routine checks. Spec-driven, approval-gated, fully auditable. | Python SSH SQLite CI |
| visor-productividad | Operational productivity dashboard over live business data, with scheduled ETL jobs syncing on-prem PostgreSQL to the cloud via systemd timers. | Next.js PostgreSQL GCP systemd |
| pedidos-materia-prima | Raw-material ordering app with authentication and role-based access. | Next.js Supabase Vercel |
| sentiment-analysis-word2vec-rnn-lstm | Comparative study of ML vs. deep learning for binary sentiment classification on IMDb 50K. | Word2Vec RNN/LSTM |
| waste-sorting-with-YOLOv8 | Computer-vision waste classification pipeline trained on a custom annotated dataset. | YOLOv8 Roboflow |
| ETL-PROJECT | End-to-end extract–transform–load solution for a real analytics problem. | Python SQL |
| Proyecto_final_ML | Interactive app predicting employee attrition with a stacking ensemble. | Streamlit scikit-learn |
Languages Python · SQL · TypeScript · Shell
AI / ML scikit-learn · TensorFlow · pandas · NumPy · spaCy · YOLOv8
Agentic LangChain · RAG · ChromaDB · MCP · LLM tool-use · eval harnesses
Cloud Google Cloud · AWS · Vercel · Supabase
Data PostgreSQL · MySQL · SQLite · Airflow · Power Query
Web Next.js · React · TailwindCSS
Ops Linux · WSL2 · systemd · Docker · Git · CI/CD · Postman
Methodology Agile · Spec-driven · MLOps · LLMOps · AgentOps
Hardware Circuit design · embedded controllers
LinkedIn · junchocaliv@gmail.com · All repositories
Observe → diagnose → report → ask → act → verify.