Junior Software / AI Developer who'd rather ship a rough prototype than read one more tutorial.
I build backend systems, ML pipelines, and small tools that solve one problem well — then push them until they actually work, not just until the demo does.
👉 adrian-pliego.vercel.app — go take a look, it doesn't bite
| 🧠 AI / ML pipelines | Forecasting, anomaly detection, and applied research — built with PyTorch and scikit-learn, evaluated honestly |
| 🔁 Automation & tooling | Python tools that replace repetitive manual work, with real safety rails instead of blind auto-pilot |
| ⚙️ Developer infrastructure | CI/CD pipelines and small open-source tools that solve exactly one annoying problem |
| 🧩 Systems & OOP design | Modeling real entities and relationships with SOLID-minded C++/Java, not defaulting to a framework because it's expected |
| 🧪 Applied experiments | Projects built to answer a specific question — including the ones where the answer wasn't flattering |
|
A curated, pre-wired setup for Claude Code that combines the best community add-ons into one install — instead of five separate repos you'd have to find, compare, and glue together yourself. Problem: Claude Code's ecosystem has a dozen good ideas scattered across a dozen repos, several solving the same problem twice.
Built: benchmarked the candidates (e.g.
Why it's interesting: zero inbound ports and an opt-in, off-by-default support nudge — the same "don't be sneaky" product instinct as |
Open-source pipeline that builds unsigned iOS Problem: testing your own iOS app normally means owning a Mac or paying Apple. Built: a GitHub Actions workflow that spins up a macOS runner to compile any Expo/React Native project, using scoped fine-grained tokens to securely check out a different target repo.
Why it's interesting: it's pure CI/infrastructure engineering — no app code, just a secure, reusable build pipeline solving a real cost problem. |
|
The retro-terminal-OS portfolio this README keeps pointing you to — boot sequence, a cursor-reactive canvas hero, and six real projects each visualized with a metaphor matched to what they actually do. Problem: a conventional Hero → About → Skills → Projects page doesn't say anything about how you build. Built: Vite + React 19 + TypeScript, hand-rolled canvas2D renderers (no WebGL) — including a hidden ASCII bust generated from a real photo via luminance + Sobel edge detection — plus GSAP-pinned scroll and Framer Motion reveals.
Why it's interesting: |
📈 Quant research — macro news forecasting · mandate investor profilingTwo-part BSc final year project: can daily macro news predict next-day ETF returns, and can investor "mandates" (not just a risk score) drive better portfolio allocation? Built: a PyTorch MLP vs. 5 classical baselines under strict walk-forward validation for the forecasting side; a Random Forest mandate predictor feeding a regime-aware, backtested ETF allocator on the portfolio side.
Why it's interesting: both repos report the results that didn't work too — e.g. the Markowitz baseline beating the mandate strategy on Sharpe — instead of only showing wins. |
|
Self-hosted pipeline that turns a script into a narrated, AI-illustrated video — entirely on local hardware. Problem: generate narrated AI-image videos end-to-end without paying for cloud inference. Built: n8n orchestrates a fully local flow — Kokoro TTS, ComfyUI (Krea2 Turbo + a style-reference LoRA) for character-consistent scene images, and an ffmpeg assembly service.
Why it's interesting: documents real hardware limits honestly — 6GB VRAM minimum, ~30GB peak RAM, ~90s per image — instead of glossing over what it actually takes to run. |
Unsupervised anomaly detection over OpenSSH logs — flagging suspicious activity without labeled attack data. Problem: in security logs, "normal" vastly outweighs "attack," and clean labels rarely exist. Built: a config-driven pipeline (Isolation Forest, LOF, One-Class SVM) with temporal feature engineering, weak-label heuristics for evaluation, and PR-AUC/Recall@K as proxy metrics.
Why it's interesting: forces careful evaluation design when ground truth barely exists — accuracy alone would be meaningless here. |
- 🔧 Actively iterating on overclaude — my most recently pushed project, open source and open to issues/PRs
- 🖥️ Rebuilt my portfolio site from scratch as ADRO_OS — React 19 + TypeScript + Vite, deployed on Vercel
- 📊 Working through the next steps I flagged myself in the FYP repos — time-series cross-validation and better regime coverage for the portfolio backtests
- 🧰 Looking for the next small, annoying manual task worth turning into a tool — that's how
auto_applyerstarted
Languages
AI / Machine Learning
Backend, automation & tooling
Frontend
CI/CD, DevOps & hosting
Practices
Primary focus: Python, PyTorch/scikit-learn, GitHub Actions · Also used: React/TypeScript (current portfolio, Vercel-deployed), Java, C++, HTML/CSS
- Prototype first, read the docs when it breaks — not before
- One command that runs the whole pipeline beats ten manual steps in a README
- Walk-forward validation isn't optional when the whole point is "did this actually generalize"
- If a project of mine has a
Limitationssection, I probably wrote it myself before anyone had to ask - Reach for OOP + SOLID when the domain has real entities and relationships to model — see
abCPLUS'sPaciente/Medico/Citadesign — not just because it's expected - Work in short, iterative loops with visible scope cuts and honest retros — Agile as a working habit, not a slide
Open to junior backend, AI/ML, and Python engineering roles — and always up for talking about a weird technical idea.
you scrolled this far — here's some ascii plasma 🌀