A document Q&A system built like a product, not a tutorial. The differentiator is the layer most "RAG chatbot" portfolio projects skip entirely: a measured retrieval eval harness, citation grounding, and hallucination detection — not just a chatbot that sounds confident.

• 📏 965 chunks over OSHA 29 CFR 1910, measured against a 45-question hand-verified eval set — every citation machine-checked against the source text
• ✅ 60 of 60 citations resolved to a real paragraph. Zero fabricated. Gold-citation rate 37/38, ungrounded numbers 0, and 7/7 out-of-scope questions declined instead of guessed at
• 🔍 The UI separates paragraphs the model cited from ones it retrieved-but-ignored, since the correct source is often sitting in that second list when the model misattributes a claim
• 🗺️ A corpus map renders all 937 indexed paragraphs as cells and lights the ten a question retrieved — the search step made visible rather than asserted
• 🔐 JWT auth with role-based access control enforced in the retrieval SQL, so a viewer is refused gated content even calling the API directly, not just in the UI
• 📊 Measured on two embedding backends — local nomic-embed-text and hosted bge-base-en-v1.5 — and the deltas are published rather than hidden

_{Demo sign-in: viewer / viewer-pass (general) or officer / officer-pass (safety officer). Ask both "Who may remove a lockout device?" to see the access gate. Free-tier hosting sleeps when idle, so a first request can take about a minute.}

An employee onboarding/offboarding provisioning system built as nine n8n workflows over PostgreSQL. A request becomes an ordered plan derived from a policy table, privileged grants pause for human approval, and every step executes idempotently and is verified by reading the target system back. The hard part was never calling the APIs — it's that APIs fail halfway, and a half-provisioned account is worse than none because nobody knows it exists.

• 🔄 Compensating-transaction saga — a failed run walks its own ledger backwards and undoes what already succeeded, in reverse order. Verified by injecting a fault into the identity API, not by reading the code
• 🔑 Idempotency at two layers — deterministic per-step Idempotency-Key headers plus a UNIQUE-constrained request ledger, making duplicate submissions impossible rather than unlikely
• 🔬 Read-back verification — a 2xx is treated as a claim, not evidence. Every write is followed by a GET that asserts the intended effect. This caught a real case where the API reported success and nothing had changed
• ⏸️ Human-in-the-loop approval that is single-use and expires in 24 hours, on a durable Wait node with the decision persisted to the database — nothing is provisioned while it waits
• 🧨 Verified by fault injection across eight failure scenarios, including killing the provider mid-rollback to prove the system distinguishes rolled_back from failed and escalates only the second
• 🤖 One tightly fenced LLM drafts requests from free-text HR email using an enum-constrained schema and evidence grounding — it produces a pre-filled form a human submits, and has no authority over anything that gets provisioned. Stripping its "quote the source" rule made it reject all seven extracted fields, which is the test most LLM-validation code never gets

_{A working demonstration, not a deployment. The identity provider is a simulator included in the repo, all employee data is synthetic, and it has never run in a company. What is real is the behaviour under failure.}

A voice-driven calendar assistant on a Raspberry Pi 5 that announces upcoming appointments unprompted and repeats until acknowledged — because a reactive assistant is useless when the failure mode is forgetting.

• 🎙️ Always-on "Hey Jarvis" wake word (openWakeWord) → faster-whisper STT → a pure, fully-tested intent parser
• 🗣️ Piper TTS voice output, half-duplex echo gating so it doesn't hear itself talk
• 💬 Optional conversational layer via Gemini — deterministic commands (acknowledge/snooze) always stay local; an LLM never decides anything that could mean a missed appointment
• ✅ Read-only Google Calendar OAuth, kept that way on purpose for an honest privacy story

A retrieval-augmented chatbot embedded in my portfolio that answers visitor questions about my work from a single curated document — and refuses rather than guesses when the document doesn't cover it. One invented credential on my own portfolio costs more than the feature is worth.

• 🌐 Runs on Cloudflare's edge through a native Workers AI binding — no API key anywhere in the deployed code
• 🧠 Embeddings via bge-base-en-v1.5, generation via Llama 3.1 8B, answers grounded only in retrieved context
• ⚡ Frequent questions have pre-written answers that cost zero inference to serve
• 🖥️ Also ships as a local app: Ollama + FastAPI + Chroma, hand-rolled retrieval loop, no LangChain

A cross-platform finance app (Android + Web, single Flutter codebase) with a built-in assistant. People track spending, set budgets and savings goals, and ask a Gemini-powered chat about their own money in plain language — every answer grounded in real transactions, never generic advice.

• Full stack shipped solo: Flutter frontend, FastAPI backend, PostgreSQL via SQLAlchemy, JWT auth
• GitHub Actions pipeline builds, signs the Android release, and deploys to Firebase Hosting + Render

A YOLOv11 + OCR pipeline that reads license plates from a camera feed, validates them against a database, and keeps a searchable record. Presented to the Land Transportation Office for accuracy approval, and successfully defended as our undergraduate thesis with Engr. Al Christian L. Kardinas.

• >92% plate-read accuracy (YOLOv11 + OCR)  ·  100% database validation accuracy

Collects, cleans, and visualizes Steam game data in an interactive dashboard, then has the Gemini API write an executive summary of what the numbers say — instead of leaving a wall of charts for the reader to interpret.

Centralizes announcements, events, student organizations, academic resources, a marketplace, lost-and-found, and messaging for TUP Visayas in one place, with a different view and permission set for students, faculty, and administrators.

• 🔌 Built against an existing production MySQL backend I wasn't allowed to redesign — the frontend types mirror the live schema and every data call routes through a single file, so going live is a one-file change
• 🎨 The design system is drawn from the campus logo itself, down to the printed-circuit traces used as an ambient background

A computer vision model that recognizes over 1,000 Pokémon species from an uploaded image, built to learn the full CV workflow rather than a ten-class tutorial version of it — and to know when an image isn't a Pokémon at all.

• 🎯 94.8% top-1 accuracy, 97.3% top-5, across 1,009 classes
• 🚫 A dedicated rejection class catches non-Pokémon images with 97.6% recall instead of confidently guessing
• 🌐 Quantized to ~7 MB and runs entirely client-side in the browser, no server involved — try it live

Commissioned by a local online bakery business to build a React/TypeScript site with data-driven content layer so the owner can add menu items without touching code, custom crossfade photo galleries built from raw phone photos, and a resilience layer that keeps animations working even inside Instagram's in-app browser, where standard scroll-trigger APIs are known to silently fail.