Languages

Machine Learning

_{Fine-tuning with reproducible seeded runs · dataset augmentation · held-out evaluation · model comparison · decision-threshold tuning · bootstrap confidence intervals · stratified k-fold CV · leakage control}

AI & Agentic Systems

Frontend & Frameworks

Backend & APIs

Databases & Data

Mobile

Cloud, Deployment & DevOps

Architecture & Practice

An AI road-condition and safety assessment system applying the international iRAP Star Rating methodology and ASTM D6433 pavement standards. I built it independently after it was turned down as a thesis proposal, then pitched it to the Toll Regulatory Board — that presentation is what opened my consultancy there.

I own the full ML lifecycle. Dataset ingestion from public research corpora (CrackForest, RDD) → geometric and appearance augmentation → YOLOv8 fine-tuning → held-out evaluation → model comparison → decision-threshold tuning → bootstrap confidence intervals.

Training is engineered for defensibility in a regulatory setting:

  fixed seeds                  reproducible runs, end to end
  SHA-256 weight provenance    identify the exact base weights, not just the filename
  per-run training records     which images and label-review states produced this model
  no-network guarantee         enforced in code and in the test suite, not by policy

A finding derived from a model whose training data cannot be reconstructed cannot be defended when a regulator is challenged on it. That constraint shaped the whole pipeline.

Six analysis modules — pavement distress & PCI scoring · roadside severity classification · IRI roughness from phone accelerometer traces via a quarter-car model · nighttime lighting assessment · KDE crash-hotspot detection · random-forest risk fusion.

~19,500 lines · 125 modules · 63 test suites  ·  Python · YOLOv8/Ultralytics · PyTorch · OpenCV · scikit-image · FastAPI · Flutter · Next.js

Built on a personal independent contract for the Toll Regulatory Board, Republic of the Philippines (2026, completed), and for private clients. Agency source and data are confidential and handled under RA 10173; details are available under NDA.

_{HopeSMS was built together with Eran Josh Reyes; the repository has the full pull-request history.}

  

Programmes and specializations (individual courses inside them are not listed separately). 45 credentials in total; every public verification link is on the portfolio.

Working through the GoF catalog in Java, one small domain model per pattern. All public:

I write the specification before the agent runs. I direct Claude Code as an orchestrated build system — written specs, custom subagents with explicit delegation boundaries, MCP server integration, project instruction sets encoding domain rules — while keeping architectural ownership and reviewing every diff.

Deterministic pre-parsing before a model call, so the model is only asked what it is actually needed for. A re-prompt-once-then-fail-loud pattern for malformed structured output, because a system that quietly accepts bad data is worse than one that stops.

I remove AI where it cannot be defended. On a regulated government system I replaced an LLM-driven evaluation layer with deterministic rules — "the model judged it this way" is not an answer an auditor accepts. AI stayed for ingestion, extraction, and drafting; nothing load-bearing for a decision.

  TypeScript / Next.js    ██████████████████░░  90%  production platform
  Claude Code / MCP       ██████████████████░░  90%  daily agentic workflow
  PostgreSQL / Drizzle    █████████████████░░░  85%  schema · rbac · audit · migrations
  Supabase / RLS          █████████████████░░░  85%  db · auth · row-level policy
  Python / FastAPI        █████████████████░░░  85%  ml pipeline + ai backends
  React + Vite            ████████████████░░░░  80%  client + internal systems
  Computer Vision         ████████████████░░░░  80%  yolov8 fine-tuning · opencv · evaluation
  Java                    ████████████████░░░░  80%  primary academic language
  Agentic Systems / RAG   ████████████████░░░░  80%  tool-calling · pgvector · guardrails
  SOLID Principles        ████████████████░░░░  80%  applied across all repos
  Tailwind CSS            ████████████████░░░░  80%  utility-first styling
  ──────────────────────────────────────────────────────────────────────
  NumPy / pandas          ██████████████░░░░░░  70%  data processing · ml pipeline
  JavaScript              ██████████████░░░░░░  70%  web + react
  Design Patterns (GoF)   ██████████████░░░░░░  70%  catalog in progress
  Zod / Pydantic          ██████████████░░░░░░  70%  typed validation layers
  Vercel / Cloudflare     ██████████████░░░░░░  70%  deployment · workers
  scikit-learn            ██████████████░░░░░░  70%  pipelines · stratified cv · roc-auc
  UML / System Design     ██████████████░░░░░░  70%  architecture documentation
  ──────────────────────────────────────────────────────────────────────
  PHP / Laravel           ████████████░░░░░░░░  60%  coursework + web apps
  Oracle Cloud            ████████████░░░░░░░░  60%  iaas · data · ai
  IBM DB2 Cloud           ████████████░░░░░░░░  60%  relational db · star schemas
  Firebase + Google Auth  ████████████░░░░░░░░  60%  authentication
  GSAP / Framer Motion    ████████████░░░░░░░░  60%  scroll-driven interfaces
  ──────────────────────────────────────────────────────────────────────
  Flutter / Dart          ██████████░░░░░░░░░░  50%  mobile capture app
  Docker                  ██████████░░░░░░░░░░  50%  containerised services
  n8n                     ██████████░░░░░░░░░░  50%  workflow automation
  Three.js / R3F          ██████████░░░░░░░░░░  50%  webgl product reveals

ico-portfolio.vercel.app  ·  ico.webdev@gmail.com

_{Open to collaborations, code reviews, and conversations about software design.}