M.S. in Artificial Intelligence · B.S. in Computer Science

I build applied AI systems across machine learning, deep learning, NLP, computer vision, and LLM-powered agents — and I'm at my best on hard, open-ended problems where the answer isn't in a textbook yet. I like to dig into the research, find what others have tried, and build my own technique when nothing off-the-shelf fits.

I've taken part in research and development programs sponsored by or partnered with John Deere, the U.S. Space Force, and UT Health San Antonio.

Open to Data Science / AI / ML Engineering roles.

Python · FastAPI · Claude · RAG · MCP · Twilio · Docker

A phone-capable AI agent that takes restaurant orders over a real phone line. It greets callers, answers menu questions from retrieved menu evidence, handles dietary/allergen questions cautiously, builds and modifies the order, tracks a running total, reads the order back, and returns a confirmation code. RAG-grounded on the menu with sentence-transformers, an MCP tool layer driving the order logic, and Twilio Programmable Voice for live calls — all reproducible via Docker Compose.

Python · MoveNet · MediaPipe · TensorFlow · Computer Vision

Analyzes weightlifting form (squat, deadlift) from a single phone video and decides whether a rep was done correctly. The hard part was monocular 3D pose estimation — recovering accurate joint angles from one camera at an arbitrary angle. It pairs MoveNet Thunder with Kalman filtering for stable tracking, auto-detects front/side/oblique views, and uses a geometry engine to "un-rotate" oblique footage and measure true squat depth, knee valgus, and hip flexion — then segments the eccentric/concentric phases to score the lift.

Python · Monte-Carlo Tree Search · Alpha-Beta Pruning · Deep Learning · Bitboards

A Reversi-playing agent combining Monte-Carlo Tree Search, alpha-beta pruning, and an AlphaZero-style deep-learning player. It took first place in both competitions in my graduate AI course. The key performance unlock was a bitboard board representation (my contribution) — the board encoded as 64-bit integers with bitwise move generation, JIT-compiled with Numba — which made search fast enough to reach far greater depth within the per-move time limit. (Team project.)

Python · NLP · Transformer Embeddings · PyTorch · Geospatial ML

A research project (UTSA) that predicts the geographic coordinates of a tweet from its text alone — no profile metadata. Transformer embeddings feed a regression head that maps language directly to latitude/longitude. We assembled an 11-million-tweet dataset (real geotagged tweets plus LLM-synthesized, news-grounded tweets) and reached a 754 km mean Haversine error — beating published baselines (PSOP, UnicodeCNN) by 2–4×, with a downstream application flagging PFAS-contamination zones. (Team research with B. Hinkley & J. De La Rosa.)

Languages: Python · Java · SQL · JavaScript · Dart · Bash

ML / AI: Deep Learning · Natural Language Processing · Computer Vision · Reinforcement Learning · Retrieval-Augmented Generation (RAG) · LLM Agents

Libraries & Tools: PyTorch · TensorFlow · scikit-learn · MediaPipe · sentence-transformers · NumPy · pandas

Backend & Data: FastAPI · Streamlit · PostgreSQL · Docker · MCP

• LinkedIn: fernando-canseco
• Email: FernCanCode@gmail.com