HL
Portrait · Ref 001
Profile
DS-0417 · rev 2026.08
Based
Cosenza, Italy · UTC/GMT +2
Field
Applied ML & AI systems
Status
Open to selected work

Héctor
Licea

Data Scientist—AI & ML Systems—MLOps

I built this page with a simple goal: to show, rather than tell, my skills and the projects I've worked on in AI, Machine Learning and beyond.

The idea comes from a difficulty I've run into many times: explaining what I can do without turning everything into a long list of technologies, skills and keywords. Sometimes words aren't enough; other times they even risk making everything look more complicated than it really is.

That's why I decided to build an interactive portfolio, where you can actually see what I've done, how I did it and, when possible, why I chose to do it that way.

So think of it as a kind of personal lab: a place to collect the projects, experiments, ideas and skills I've developed along the way. It's not a definitive snapshot, but rather a work in progress of what I can do so far.

There is, however, one thing I consider more important than any skill or technology: trying to understand how the underlying mechanisms work. It's an approach that takes time and often means going much deeper than necessary… but once you understand the mechanism, many of the ideas for implementing it come almost on their own.

I hope this page manages to tell you a bit more about who I am, how I think and, above all, what I love to build.

Technical domainsHover to isolate cluster
Embedding space — live projectionStreaming
UMAP · DIM 2UMAP · DIM 1
23
Models in prod
41
Pipelines in prod
1,284
Experiments logged
10yr
Shipping ML
P.S. 1

I'm not a frontend developer. I don't enjoy doing frontend and, with all due respect to those who do, I have no intention of becoming one. 😄 Claude and OpenAI helped me solely with connecting and polishing the frontend pieces. The backend, the logic, the algorithms, the ML and the AI are all my own work. In short: they make the interface speak nicely, I take care of everything that happens behind it.

P.S. 2

What you see in the background isn't just a visual effect. It's a K-Means algorithm running in real time, which I developed in Python. To choose the optimal number of clusters I use the Silhouette Score. In short, even the background has a job to do.

02

Projects

Filter by stack
P-01

UNICAL Games

Deployed
Problem
A drawing-guess game needs scoring that can't be gamed by just drawing more — and it needs to feel instant, not batch-graded after the fact.
Solution
Each submitted canvas runs through a YOLOv8n object detector at inference time; the score is built directly from the returned labels and confidences.
YOLOv8UltralyticsFastAPINext.jsDocker
Open project
P-02

UNICAL GAMES Dataset Builder

Deployed
AutoLabel4 of 10 gallery drawings
circlehousecarbicycle
Problem
An object detector needs thousands of boxed examples per class, and hand-labelling doodles is slow — but Quick, Draw! ships strokes, not images or boxes.
Solution
Render each stroke sequence to an image, find the ink's extreme pixels to box it automatically, and write the box as a YOLO label — then compose multi-object scenes from the results.
PythonOpenCVmatplotlibQuick, Draw!YOLOCOCOReactCanvas
10,000
Single-object images
1,500
Composite scenes
0 px
Box error vs notebook (gallery)
Open project
P-03

Plate Tracker

Deployed
Detections replayed from a 1,800-frame exportavg confidence 0.54
0.760.54
Problem
Demoing a detector live in a browser means paying inference cost on every visit and keeping a model server up just to show a portfolio piece.
Solution
Run YOLOv8 with tracking once, offline, over a real test clip; export every frame's boxes, confidence and track id to JSON; replay that JSON client-side, synced to the looping video by timestamp.
YOLOv8UltralyticsByteTrackReactSVG
1,800
Frames processed
83.7%
Frames with a detection
2,052
Boxes replayed
Open project
P-04

Corso Mazzini Dataset Constructor

In progress
From a live webcam to a labelled datasetAirflow · Selenium · Label Studio · YOLO
retrain every 100 → pre-labelWebcam1-min clipsFramesLabel StudioYOLO
Problem
A detector for one specific street needs pictures of that street — its angle, its light, its crowds at every hour — and no public dataset has them. The only source is a public webcam that streams live and keeps nothing.
Solution
Schedule the capture as an Airflow DAG: Selenium brings the webcam's player to a live state, the stream is recorded in one-minute chunks, the chunks become frames, and the frames are labelled in Label Studio with a YOLO model that is retrained every 100 labels to pre-annotate the rest.
Apache AirflowSeleniumFFmpegLabel StudioYOLOPython
60s
Recording chunk
1,440
Chunks per 24 h run
100
Labels per retrain
Open project
P-05

Corso Mazzini

Deployed
One frame's boxes, and where four of those people walked31,568 detections · 324 ids
homography
Problem
A detector's output is a stack of boxes in image space. Nobody can read footfall, dwell or desire lines off boxes — the question a square actually poses is where people walk, and that lives on the ground, not on the sensor.
Solution
Detect and track once, offline; take each box's ground point; push it through a homography onto an orthophoto of the square; ship the recorded result and replay it against the source clip, so both coordinate systems are visible at the same instant.
YOLOObject trackingHomographyOpenCVReactCanvas
31,568
Detections replayed
324
Track identities
14.6
People per frame (avg)
Open project
P-06

School Psychologist Allocation

Deployed
Cosenza · 15 psychologists244 schools → 68 groups
Problem
A fixed number of psychologists has to cover 17,709 students in 244 schools spread over a mountainous province. Each psychologist can follow only so many students, and every school assigned far from where they are based costs hours on the road.
Solution
Reduce first, then optimise: K-means folds the 244 schools into 68 nearby groups, psychologists are placed where the students are, and an integer program assigns whole groups to psychologists, minimising travel without exceeding capacity — or, when there are not enough psychologists, spreading the overload as evenly as possible.
Pythonscikit-learnSCIPPySCIPOptLeafletReact
244 → 68
Schools reduced to groups
0
Precomputed scenarios
8.9 km
Average trip, 15 × 1,300
Open project
P-07

Real Estate Comparables Finder

Deployed
South Florida, Apr 2026Q1 listings vs. sales
Problem
An investor scanning hundreds of active listings cannot price each one by hand. A low list price alone says nothing: it may be a bargain, or simply a smaller, older home in a cheaper street.
Solution
Do what an appraiser does, automatically: for every cheap listing, find the homes sold nearby, keep the ten most similar in age, size and construction, and compare the asking price with what those homes actually sold for.
Pythonpandasscikit-learnFoliumopenpyxlLeafletReact
0
Active listings screened
0
Opportunities ≥ $120k below comparables
−34%
Median gap to comparable sales
Open project
P-08

PP-OCR Autopsy

Deployed
PP-OCR, taken apartdet → crop → CTC
« Analisi dei dati »
Problem
An OCR library is a black box. When it returns garbage (a symbol for an icon, two lines fused, a table cell gone) there is nothing to fix unless you know which piece failed and why.
Solution
Run every component in isolation, record what goes in and what comes out, and rewrite each non-neural step until it gives exactly PaddleX's result. What is left is a few networks wrapped in ordinary code.
PaddleOCRPP-StructureV3OpenCVNumPyPythonTypeScriptReactDocker
0
Text lines traced end to end
217/217
Lines identical to PaddleOCR
6.6 MB
Networks behind detection + recognition
Open project
03

Experience

2023 — Present3 yr · current

Lead Data Scientist — AI Systems & Models

WishInnovation/Cosenza/Design, development and deployment of AI systems and models (and plenty more)
  • Own the applied ML and AI roadmap.
  • Built end-to-end AI systems for the company.
  • Led the adoption of new technologies and methodologies to make AI projects more efficient.
  • Developed and deployed production-ready, scalable AI models.
  • Developed end-to-end pipelines for training and deploying AI models.
  • Built agentic systems using both on-premise and cloud AI models.
~4AI products1AI Literacy framework1AI adoption strategy (no AI Slope)
Stack  Python · Docker · Apache Airflow · Power BI · Label Studio · OpenCV · Open Linked Data · SPARQL · Neo4j · Cypher · Knowledge Graphs · PostgreSQL · SQL Query Language · SDMX (European Statistical Data and Metadata Exchange) · Ollama · vLLM · OpenAi Models · LangGraph · LangChain · Web Scraping · Scrapy · API Integration · FastAPI
06/2025 - 08/20253 mo

Python Lecturer

UNICAL (DISCAG)/Cosenza/Master's in Green Transition and Climate Risk Management
  • Taught hands-on Python lessons to the master's students.
  • Prepared a detailed syllabus for the course.
  • Designed practical exercises and teaching materials to support student learning.
  • Prepared and graded student exams and projects, giving detailed feedback to improve learning.
30Hours of teaching1Python for Data Science course30hHands-on experience teaching Python in Italian
Stack  Python · Pandas · NumPy · Matplotlib · Jupyter Notebook
2018 — 20213 yr

Instructor Professor — Cuban Economy

Universidad de Oriente/Cuba/Cuban Economic History & Economic Models and Econometrics
  • Taught courses on the Cuban economy and economic models, developing teaching materials and supervising student research projects.
  • Prepared and graded student exams and projects, giving detailed feedback to improve learning.
  • Prepared teaching materials and resources for student labs and practical sessions.
1Publication
Stack  R · Excel · Stata · SPSS
04

Capability map

Domain dependency graphHover or focus a node
Data EngineeringData ScienceMachine LearningKnowledge GraphsComputer VisionLLM SystemsMLOpsAI Systems
CORESTRONGWORKING
Domain detailNODE 03 / 08

Machine Learning

The centre of gravity. Supervised learning on tabular and sequential data, with more time spent on evaluation design than on architecture choice.

Stack
PyTorchCORE
LightGBMCORE
ForecastingCORE
Evaluation designCORE
Bayesian methodsWORKING
10 YR · 23 MODELS IN PRODUCTION
05

Lab notebook

EXP-041

Steering with sparse features

Running

Hypothesis — Features recovered by a sparse autoencoder on mid-layer activations can steer refusal behaviour more precisely than prompt instructions, without collapsing answer quality.

Steered refusal +31%Quality delta −2.1%n 1,800
InterpretabilitySAESteering
EXP-038

Quantisation vs. retrieval recall

Concluded

Hypothesis — Binary quantisation of embeddings costs less recall than the memory saving suggests, once a small float rescoring pass is added on top.

Memory −96%Recall@10 −1.4%Rescore top-200
RetrievalQuantisationAblation
EXP-034

Graph-conditioned RAG

Running

Hypothesis — Conditioning retrieval on an entity graph beats flat chunk similarity on multi-hop questions, where the answer is never in one passage.

Multi-hop EM +17ppSingle-hop no changeLatency +180ms
RAGKnowledge graphsMulti-hop
EXP-029

Synthetic minority defects

Parked

Hypothesis — Diffusion-generated defect crops improve recall on classes with fewer than 200 real examples. Result — they help until roughly 400 real samples exist, then actively hurt precision.

Recall, rare classes +9ppPrecision −4ppVerdict conditional
Synthetic dataImbalanceVision
Written & released5
Reconciled hierarchical forecasting in a live supply chainINDUSTRY TRACK · 2023 · PDF
Edge-distilled segmentation under a 16ms frame budgetWORKSHOP · 2021 · PDF
Representation learning for industrial sensor streamsWORKSHOP · 2017 · PDF
KELVIN-SURF — 41k labelled defect crops, 3 classesDATASET · CC BY-NC · 2019
tinyrerank — cross-encoder reranker, 22M paramsMODEL · APACHE-2.0 · 2024
Currently readingQ3
Evaluation design for generative systemsREADING GROUP · WEEKLY
Causal inference for observational product dataSELF-STUDY
Mechanistic interpretability — sparse dictionariesFEEDS EXP-041
06

Contact

>whoami
héctor licea — human being, data scientist, ai & ml systems
>availability --status
open to selected consulting and lead roles
Open a thread→
LOCATION
Cosenza, Italy · UTC/GMT +2
LANGUAGES
Spanish (native) · English (professional) · Italian (professional)
REMOTE
Yes — EU timezones preferred
© 2026 Héctor LiceaDesign concept · placeholder contentBuild 2026.08Rendered client-side · no tracking