Backend & Machine Learning Engineer

High-concurrency bet validation in Rust

Accepts live betting tickets and validates odds, checks balance and debits it inside a single Redis Lua script, then acknowledges the client before the write reaches PostgreSQL. Persistence runs off the request path through a Redis Streams consumer group that replays its pending list on startup, so a worker killed mid-batch re-processes rather than drops. Money is i64 cents end to end — the domain layer has no floating-point arithmetic.

• 1,000 req/s at p99 300 ms under k6 load tests
• Hexagonal layering: domain declares ports as traits, imports no infrastructure
• Unlabelled Prometheus counters, so series count cannot grow unbounded

• Rust
• Actix-Web
• Redis Streams
• PostgreSQL
• Docker
• k6
• Prometheus

Source code ↗

Token identity with refresh-token theft detection

Refresh tokens are opaque, stored only as SHA-256 digests, and grouped into families where each token is exchangeable exactly once. That single-use rule turns theft into a detectable event: if a token is ever exchanged twice, the service revokes the whole family instead of guessing which side is the attacker. Access tokens carry jti and ver claims, so a live JWT can be withdrawn before it expires.

• Rotation race closed in SQL (UPDATE ... WHERE used_at IS NULL), not in application code
• Redis deny list keyed by jti with TTL equal to the token's remaining life
• Per-endpoint permission RBAC; one version bump retires unbounded live tokens

• Rust
• Axum
• PostgreSQL
• Redis
• JWT
• SHA-256

Source code ↗

Rust/Axum backend for a lead-generation catalogue

A catalogue platform for companies that do not sell online but generate leads through quotations. The public side browses products with search, filters and PDF datasheets and submits quote requests; behind it sits an admin panel with full CRUD over products and categories, file uploads and request management. Peruvian tax IDs are validated with the Módulo 11 checksum rather than a length check, so an invalid RUC is rejected at the form instead of downstream.

• Rust and Axum over PostgreSQL, with S3-compatible file storage
• JWT authentication with Argon2id password hashing
• Astro and TypeScript frontend with Nano Stores for shared state

• Rust
• Axum
• PostgreSQL
• JWT
• Argon2id
• Astro
• TypeScript
• S3

Source code ↗

Football retrieval and ranking over 1.75M event actions

Ingests match records, rosters and event streams into a relational schema, compresses player-seasons into PCA embeddings, and serves item-item similarity search: given a player-season, it ranks the closest comparables from a position-filtered pool. The same embedding backs a clustering layer, a k-NN similarity graph and an ILP starting-XI optimizer.

• 94,525 matches · 1,751,751 event actions · 52,387 player-seasons
• Recall@10 0.1435 vs 0.0614 baseline; MRR 0.1788 vs 0.0773; NDCG@10 0.1146 vs 0.0430
• Leave-one-out over 3,670 player-seasons, 1,662 queries, identical evaluation pool for both systems

• Python
• Polars
• pandas
• scikit-learn
• PCA
• Parquet
• Streamlit

Source code ↗ Live demo ↗

Distributed gradient descent in Go over 200,000 records

Three generalized linear models — one logistic, two linear — trained by gradient descent in Go to predict mortality risk, survival days and treatment cost. Training parallelizes two ways over the same map-reduce code: a goroutine fan-out inside a node, and a parameter-server cluster over net/rpc across nodes. The REST, JWT and WebSocket layers are standard library; go.mod declares two direct dependencies.

• Mortality AUC 0.770 against a 0.500 random-ranking reference; accuracy 0.832
• Survival R² 0.906 (RMSE 199 days) · Cost R² 0.983 (RMSE $1,494)
• 2.95× speedup saturating at 4 workers on a 4-core machine, measured over 3 repetitions

• Go
• MongoDB
• Redis
• net/rpc
• Docker

Source code ↗

Entity resolution linking candidates to public sanction records

A medallion pipeline in Polars that consolidates Peruvian presidential and congressional candidates from JNE, Congress and government transparency portals, links each one to companies sanctioned by OSCE, and scores a reproducible risk index served by a FastAPI read layer over Parquet. Deterministic ID-to-tax-ID matches and fuzzy name matches are tagged separately on every row, so an unverifiable guess never looks like a documented link.

• Five-stage layout: raw → staging → normalized → matched → curated
• Financial risk on median + MAD × 1.4826 rather than mean and standard deviation, because declared assets are heavily skewed
• No labelled ground truth for the matcher — precision and recall are unmeasured, and the repository says so

• Python
• Polars
• FastAPI
• Parquet
• FAISS
• Docker

Source code ↗

Embedding-based student/offer matching across four services

Student profiles and company offers are turned into field-weighted text, embedded with a multilingual sentence-transformer, indexed in Qdrant and ranked by cosine k-NN. A two-sided swipe layer records intent and promotes a pair to a match only when both sides swipe right. Field importance is expressed by repeating terms in the input text — required skills ×10, area ×8 — because the encoder has no per-field weighting input.

• Four services; 20 relational tables under async SQLAlchemy, laid out as ports and adapters
• Spanish text lemmatized and stop-word filtered with spaCy before encoding
• Recommender is unevaluated: no held-out split and no relevance labels yet, stated in the repository

• Python
• FastAPI
• PostgreSQL
• Qdrant
• sentence-transformers
• spaCy

Source code ↗

Licence-plate detection from detector to field apps

A YOLOv8 detector trained on 2,620 annotated Peruvian licence-plate photos, wrapped in a Flask service that crops the plate and runs OCR over it. An Electron desktop app is the inspection station and an Expo mobile app is the field client; both talk to the same HTTP API over one Supabase schema.

• 4,144 boxes over 2,433 training images — 1.70 plates per image, so multi-object rather than single-object
• Median box 504 px², 0.12% of the frame: below the COCO small-object threshold, which is what drives imgsz
• Detector has no committed evaluation — no mAP figure is claimed, and the repository says why

Collaborative repository — hosted under a teammate's account.

• Python
• YOLOv8
• PaddleOCR
• Flask
• Supabase
• Electron
• React Native

Source code ↗

Pictogram-based communication assistant

An augmentative-communication web app that turns Spanish text into pictograms in real time. A FastAPI backend serves both REST and WebSocket chat, lemmatises incoming text with spaCy before looking the terms up against the ARASAAC pictogram set, and drives a sentence builder and a tutor assistant on a React client.

• FastAPI over REST and WebSockets; React and Vite client
• spaCy es_core_news_lg lemmatises Spanish before the pictogram lookup, so inflected forms resolve to one symbol
• Photo-to-pictogram classification is listed as conditional in the repository — treat that piece as partial

• Python
• FastAPI
• WebSockets
• spaCy
• TensorFlow/Keras
• React
• Vite
• Tailwind CSS

Source code ↗

Concurrent anonymization pipeline for judicial case files

A Go worker-pool pipeline that normalizes and anonymizes the text of Peruvian Constitutional Court case records, fed by a Python scraper and EDA layer that build the corpus. The measurement question the repository answers is how a channel-fed goroutine pool scales from 15 to 150 workers.

• 149,387 source case records; pipeline scales to a combined corpus of roughly 1.4M rows
• 10.26× throughput going from 15 to 150 workers, measured over 1,100 timed runs committed to the repo
• Efficiency above 100% is an artifact of a simulated per-record cost, not real CPU work — flagged as a limitation

• Go
• Python
• pandas
• semantic search

Source code ↗

CV ingestion, simulated AI interview, automated ranking

A recruiter-facing platform that stores uploaded CVs, extracts structured information from the PDFs, runs a simulated interview through a conversational assistant, and produces an automated ranking of candidates per job offer. FastAPI backend with a React and TypeScript client.

• Prototype stage: persistence is an in-memory cache, not a database
• Scoring combines interview performance with fit against the offer

• Python
• FastAPI
• React
• TypeScript
• Tailwind CSS
• PDF extraction

Source code ↗

Academic tutoring platform on Java + Spring Boot

A tutoring marketplace covering registration, tutor availability, session booking, simulated payments, session links and reviews. Java/Spring Boot REST backend against a relational schema, with a React and TypeScript client; the repository carries the UML model, the database diagram and the full delivered backlog.

• 16 user stories delivered end to end
• Roughly 90% of the API endpoints built by me
• Relational schema and UML diagrams committed alongside the code

Developer on a five-person university team; I built roughly 90% of the API endpoints.

• Java
• Spring Boot
• React
• TypeScript
• SQL

Source code ↗

Software Engineering Intern · IT consultancy · current role

Software engineering intern at an IT consultancy, working on internal platform and automation work. Client deliverables and internal products are confidential and are deliberately not described here.

• Current role
• Backend and automation work under confidentiality

• Python
• TypeScript
• PostgreSQL
• Docker
• CI/CD

Company on LinkedIn ↗

Co-founder · CV-matching marketplace shipped to real users

Co-founded a platform connecting university students with real project challenges posted by companies, and took it to real users. The backend owns the relational domain — students, companies, offers, skills, matches and agreements — over async SQLAlchemy and PostgreSQL, and delegates candidate ranking to a separate AI service over HTTP.

• 15 domain entities and 8 use cases behind explicit port interfaces
• 8 HTTP endpoints; FastAPI confined to the input adapter
• No automated tests and no CI test stage — stated in the repository

• Python
• FastAPI
• SQLAlchemy
• PostgreSQL
• JWT
• gunicorn
• Azure

Company on LinkedIn ↗ Source code ↗

No-Code Developer Intern · MVP delivery on Bubble

Built functional MVPs on Bubble so client ideas could be validated with users before any engineering budget was committed.

• No-code delivery on Bubble

• Bubble
• No-code

Company on LinkedIn ↗