Medellín, Colombia · remote-friendly

Alexey Velásquez.

AI engineer. I build AI that runs where your data lives: speech intelligence, grounded LLM agents, interpretable ML.

up to 80% less manual review · up to 99.5% fault detection · 7 systems on GitHub

Open to AI / ML engineering roles

Experience

Three years, one direction: from fault-detection ML in the lab to self-hosted speech intelligence in production, and now teaching managers to use AI. Each role links to its code.

  1. Jan 2026 – now

    Data Scientist · Luis Buys Homes

    Florida, US · remote

    • CRM automation suite in production: 3 event-driven pipelines, ~43,000 polling ops/month cut to ~3 per event.
    • Remote MCP server so the sales team asks their CRM questions through Claude — role-gated, read-only.
    • BERT sentiment over first-contact calls; PostgreSQL to Power BI dashboards the team ran their day on.
    • KPI engine on Cloudflare Workers behind every weekly report, guarded by 266 tests.
    • Lead enrichment of 21,828 leads against property records: 491 already-sold homes taken off the dial list.
    See it: CRM suite
  2. Sep 2026 – Oct 2026

    Lecturer, AI for Business · UNAC

    Medellín, Colombia · Specialization in Senior Management

    • Taught business managers to put AI to work in their own companies, with graded thesis projects.
    • Built a connected classroom: each student's Claude acts as their project advisor, grounded in the course rubrics.
    • Panelist, "The Future of Organizations in the Age of AI", CIEACE 2026 congress.
    See it: Aula IA UNAC
  3. Jun 2025 – Dec 2025

    Data Scientist · Cactus Home Offer

    Florida, US · remote

    • Speech intelligence for sales calls: ASR, pattern detection, SHAP. Up to 80% less manual review.
    • LLM KPI extraction in mixed Spanish and English, served in real time through n8n.
    See it: ORCA
  4. Aug 2023 – Aug 2025

    Machine Learning Engineer · EIA University

    Medellín, Colombia

    • Turbine fault detection from vibration data: LightGBM and XGBoost, up to 99.5% accuracy.
    • Model tournaments — Random Forest, SVM, autoencoders — on a PostgreSQL-fed pipeline; the winners shipped.
    See it: Turbine FDD

Studies

The formal foundation under the practice.

  1. Aug 2025

    Master's in Engineering, ML focus · EIA University

    1st Class · 94%

  2. Mar 2023

    Bachelor of Civil Engineering · Universidad de Medellín

    Licensed engineer · COPNIA

Selected work

Built end to end. Run for real.

Machine learning applied to sales-call intelligence.

A raw sales call goes in. Out come the transcript, the speaker emotions, the KPIs, a coaching scorecard and a searchable archive you can chat with. Nothing leaves the machine.

  • Up to 80% less manual call review for a US sales team
  • ~95% retrieval quality at 20% of the storage
  • PII redacted before anything is stored — 5 GPU services, one docker compose up
  • Whisper
  • pyannote
  • emotion2vec
  • FastAPI
  • pgvector
  • Ollama
  • n8n
  • Docker
  • CUDA
GitHub See it running

Web scraping enhanced with local LLMs.

Plain English in, spreadsheet out. A local LLM plans the crawl, and every extracted value is checked against the page before export.

  • Every value verified against the source page — ungrounded fields are blanked, not guessed
  • Schema-constrained decoding on every LLM call
  • 151 tests, fully offline · live crawl UI with plan, progress, ETA
  • Ollama
  • Flask
  • SQLite
  • Docker
  • pytest
  • JSON-LD
GitHub See it running

Automated workflows in Make interconnecting CRM systems.

Leads, calls and stage changes flow between the CRM, the phone system, Slack and Google Forms with no human in the middle. Ran in production for a US real estate team.

  • ~43,000 idle polling ops/month cut to ~3 per event
  • Every lead logged, deduped and routed with zero manual touch
  • A production race condition found and engineered out — authored as code, not clicks
  • Make.com
  • n8n
  • Python
  • Node.js
  • Slack Block Kit
  • Apps Script
  • Webhooks
GitHub

MCP server connecting CRM KPIs to Claude AI.

The sales team's CRM data — KPIs, lead activity, call performance — available inside Claude AI through a remote MCP server, so Claude can analyze it on demand and hand back clean HTML reports. A Cloudflare Worker maps each Google sign-in to a CRM identity and scopes every query server-side.

  • In production for a US real-estate sales team — onboarding is 1 click
  • 12 MCP tools, role-gated: closers only ever see their own data
  • 100% read-only by construction
  • TypeScript
  • Cloudflare Workers
  • MCP
  • OAuth 2.0
  • Durable Objects
  • Zod
GitHub

Training machine learning algorithms for wind-turbine damage detection.

Six accelerometers on a real 14-meter turbine in the Andes. Damage physically induced, natural frequencies distilled from raw vibration, and a five-family model tournament — with SHAP confirming the models trust the same physics an engineer would.

  • 0.973 AUC detecting damage on hold-out
  • 6,000+ vibration records over 1.5 years of field campaigns
  • The first vibration mode drops by half when damage appears — the models found it
  • PyCaret
  • LightGBM
  • scikit-learn
  • SHAP
  • SciPy
  • Jupyter
GitHub

Deep residual networks for pattern detection on raw vibration signals.

A published deep-learning method reproduced in PyTorch and trained directly on raw accelerometer waveforms from the same 14-meter turbine — then stress-tested with leakage-safe evaluation and physics-driven ablations.

  • Detection of never-seen damage lifted 47% → 98% by per-window normalization
  • Leakage-safe record-level evaluation — paper-style numbers drop 4–8 points under the honest split
  • Engineered modal features + LightGBM still win by ~20 accuracy points — and the study shows why
  • PyTorch
  • CUDA
  • Optuna
  • NumPy
  • h5py
  • pandas
Code private — under peer review for a scientific publication

Training LightGBM models to learn expert judgment in structural risk screening.

Rural Colombian schools, expert-rated for earthquake, landslide, flood and windstorm risk. Hazard-specific encodings, an 11-algorithm tournament, and a Streamlit screener that explains every rating with SHAP.

  • Up to 0.977 hold-out accuracy across four hazards
  • Every miss is adjacent-class — no Low ever rated High, no High ever rated Low
  • A scenario engine quantifies which retrofit collapses the High-risk share
  • LightGBM
  • Extra Trees
  • SHAP
  • Streamlit
  • scikit-learn
  • pytest
GitHub

BERT-based sentiment analysis of citizen service complaints.

Real citizen complaints from a Colombian aqueduct program. PII stripped with a zero-leak audit, scored by Spanish transformers running locally, and cross-examined with statistics: feeling follows process, not place.

  • 68% of complaints negative — not one positive in the whole corpus
  • Domain shift documented: a social-media emotion model abstains on 92%
  • Ships as a desktop Studio app — nothing leaves the machine
  • Transformers
  • RoBERTuito
  • PyTorch
  • pandas
  • CustomTkinter
  • pytest
GitHub

Tech stack

Hover a technology to see where it earned its place.

Core

Machine learning

Speech & language AI

Data

Infrastructure

Methods

◆ nothing selected yet

About

I'm Alexey David Velásquez Betancurt, based in Medellín, Colombia. Master's in Engineering, ML focus, 1st Class (94%). Three years building and shipping ML — from turbine sensors in a university lab to speech intelligence running in production for US companies.

I like owning the whole system, data to deploy. If it can run on your own hardware, I will make it run there. Spanish and English.

Let's talk.

Looking for an AI engineer who ships whole systems? My inbox is open.

alexeyvb135@gmail.com