Elias Arellano Campos — AI Engineer

PROJECT 01 — FLAGSHIP

Deep Learning · Healthcare

EEG Seizure Detection

Four architectures benchmarked for seizure-onset detection on 23-channel pediatric EEG — measuring exactly what temporal modeling adds.

PyTorchMNECHB-MITCNN · GRU · LSTM · TCNSaliency maps

View repository ↗

Dataset · CHB-MIT scalp EEG (PhysioNet)

23

Pediatric patients

~1,000h

Recorded EEG

198

Labeled seizures

23ch

256 Hz · EDF

Raw signal · 23-channel scalp EEG during a seizure event

23-channel bipolar-montage scalp EEG showing rhythmic, high-amplitude ictal discharge across channels FP1–F7 through CZ–PZ over roughly a ten-second window — **CHB-MIT scalp EEG** — bipolar montage, 23 channels @ 256 Hz · the rhythmic high-amplitude region is the seizure the models learn to flag · Shoeb 2009, PhysioNet

Inference pipeline

InputEEG · 23ch / 256Hz

→

Segment10s windows

→

SpatialCNN encoder

→

TemporalGRU

→

DecisionSeizure / Clear

The hard part · class imbalance

64,240 normal windows

149 seizure windowsjust 0.23% positive

64,389 totalhandled w/ weighted loss

Architecture benchmark · F1 score

CNN + GRU0.854

CNN + LSTM0.802

TCN0.722

CNN baseline0.608

Multi-patient test split (patients 1·2·3·5). CNN+GRU won — F1 0.854, AUC 0.9924, recall 0.814. Both temporal models beat the CNN baseline, confirming sequential modeling adds real value.

PROJECT 02

Machine Learning · Fintech

Fraud Detection System

Random Forest tuned for high recall with near-zero false alarms — every decision SHAP-explainable.

scikit-learnRandom ForestSHAPpandas

View repository ↗

Confusion matrix · fraud class

Predicted fraud

Predicted legit

Actual fraud

0True positive · caught

37False neg · missed

Actual legit

84False pos · false alarm

✓True neg · cleared

Holdout test set. 2,016 fraud caught vs only 84 false alarms; FN inferred from 98.2% recall. TN = the large majority of legitimate transactions, correctly cleared.

Separability · ROC-AUC

0.999

ROC-AUC

Near-perfect separation between fraud and legitimate transactions across all thresholds.

Precision vs recall · fraud class

Recall98.2%

Precision96.0%

Tuned to catch nearly all fraud (recall) while keeping false alarms low (precision) — the right tradeoff for a fraud screen.

97.1%

F1 · fraud

98.2%

Recall

96.0%

Precision

99.55%

Accuracy

PROJECT 03

NLP · Information Retrieval

Amazon Opinion
Search Engine

Swapped Boolean keyword search for SBERT semantic retrieval over 210K+ reviews — higher precision, far less noise.

SBERTNLPFlaskInverted index

View repository ↗

Retrieval pipeline

Query"audio quality: poor"

→

CleanPreprocess

→

EmbedSBERT

→

MatchCosine top-k

→

OutputRanked reviews

Avg precision ↑ (higher better)

SBERT0.578

Lang. model0.550

Boolean0.391

Docs retrieved ↓ (lower better)

Boolean1,278

Lang. model582

SBERT216

Five real-world opinion queries · +47.7% precision (0.39 → 0.58) while cutting retrieved docs 83%. Deployed as a Flask API.

Evaluation queries · aspect → opinion

audio quality → poor wifi signal → strong mouse button → click problem gps map → useful image quality → sharp

PROJECT 04

Computer Vision · Real-time

F1 Driver
Recognition

Real-time perception: webcam → face ID → API logging → live dashboard, recognition and overlay in under 100ms.

OpenCVKNNStreamlitFlaskSQLite

View repository ↗

Live F1 driver recognition: green bounding box around the detected face, name label, and a real-time stats dossier overlay

Live capture — face detected → driver identified → stats dossier rendered, <100ms

Pipeline architecture

CaptureWebcam

→

DetectHaar cascade

→

ClassifyKNN face ID

→

LogFlask → SQLite

→

VisualizeStreamlit

<100ms

Detect → overlay

10

Drivers

0–100

Fantasy score

Driver roster · trained classes

VerstappenHamiltonLeclerc PiastriNorrisAntonelli AlonsoSainzAlbonRussell

Fantasy scoring engine · weighted factors

WinsPodiumsPoints / race Win → podium conv.Pole → win conv. Fastest lapsDNF rate penalty

PROJECT 05

GenAI · Retrieval-Augmented

Tennis Coach
RAG Agent

A beginner-friendly tennis assistant that answers from uploaded coaching PDFs — TF-IDF retrieval feeding Gemini for cited, grounded answers.

Gemini APIStreamlitpdf.jsTF-IDFRAG

View repository ↗

AI Tennis RAG agent UI: sidebar with API key, uploaded tennis PDFs and a knowledge base of 3 docs and 117 chunks; the chat answers beginner questions with source-cited passages and refuses an out-of-scope question about Max Verstappen

AI Tennis — the deployed agent grounds answers in uploaded docs (source-cited, e.g. “6 source chunks retrieved”) and correctly refuses off-topic questions like “Who is Max Verstappen?”

Retrieval-augmented generation flow

SourcePDF / text

→

SplitOverlapping chunks

→

IndexTF-IDF

→

RetrieveCosine top-k

→

GenerateGemini

→

AnswerCited + grounded

Overlapping chunks prevent context fragmentation; answers carry source attribution back to the retrieved passages. The same pattern powers the live agent below.

PDF + text ingestion Overlapping chunks TF-IDF + cosine Similarity scoring Source-attributed answers

I build AI systems
that ship —
and I measure
everything.

Five systems, built end-to-end
and held to the numbers.

EEG Seizure Detection

Fraud Detection System

Amazon Opinion
Search Engine

F1 Driver
Recognition

Tennis Coach
RAG Agent

Don't read about my GenAI work.
Talk to it.

A full-stack ML toolkit,
from architecture to API.

Five systems, built end-to-endand held to the numbers.

EEG Seizure Detection

Fraud Detection System

Amazon OpinionSearch Engine

F1 DriverRecognition

Tennis CoachRAG Agent

Don't read about my GenAI work.Talk to it.

A full-stack ML toolkit,from architecture to API.

Five systems, built end-to-end
and held to the numbers.

Amazon Opinion
Search Engine

F1 Driver
Recognition

Tennis Coach
RAG Agent

Don't read about my GenAI work.
Talk to it.

A full-stack ML toolkit,
from architecture to API.