Me (Franz Eliezer Samilo), seen from behind in a black shirt, looking out over a lit city plaza at night
FrontispieceMe, after hours

A portfolio in six parts

Full-Stack Design Engineer

I build agentic AI — LLM orchestration, MCP servers, autonomous agent pipelines — and the full-stack systems that put it in front of people.

  • Lead Developer · XzenTech
  • Agentic AI · OZ Tech
  • Antique, Philippines · UTC+8

Open to remote & Philippine roles

01In my own words

Preface

I start where the logic breaks, and I build until it holds. Most of my work lives where AI meets product: agent pipelines that call real tools, LLM features that have to survive production, and the interfaces people actually touch.

Right now I lead development at XzenTech Solutions, where I've shipped six production web and mobile apps, and I build agent workflows for OZ Tech's clients. Before that I built the frontend of Rare's EarthRanger dashboard, used across 750+ conservation sites.

I'm from Sibalom, Antique — no local mentors, no startup ecosystem. Just internet access and the conviction that systems should work better than they do. I still work from there, on Philippine time, with teams in Europe, the US and at home.

I design in code. This site comes in two editions because I think both ways of reading are fair.

Based
Sibalom, Antique, PH · UTC+8
Working
Remote · open to PH onsite & hybrid
Education
BS Software Engineering, Central Philippine University (2022–2026)
Certified
6 Anthropic Academy courses — MCP, Agent SDK, Claude Code
Languages
English, Filipino
Me, taking a mirror selfie in an all-black outfit, phone partly covering my face
Self-portrait, in the usual black.Plate II
02Where I've worked

The Line

  1. Nov 2025 — PresentRemote · ContractNow

    Lead Developer

    XzenTech Solutions

    I lead development for the product suite — I've designed and shipped six production web and mobile apps, several of them with machine learning inside.

    • Cartman PH
    • XZEN Connect
    • XTS Web
    • RoboGuide
    • Proxima LMS
    • StatSmart
  2. Apr 2025 — PresentRemote · Part-timeNow

    Full-Stack Engineer — Agentic AI

    OZ Tech (unwiz.ai)

    I build agentic AI workflows for a diverse client base — LLM orchestration, MCP servers, tool chains and end-to-end agent pipelines, with GoHighLevel automation.

    • Mentoria (Greece)
    • MCP servers
    • Agent pipelines
  3. May — Aug 2025Remote · Contract

    Frontend Developer

    Rare.org

    I built the frontend for the EarthRanger conservation-analytics dashboard, now deployed across 750+ conservation sites tracking 23,000+ animals.

    • Real-time telemetry
    • GIS mapping
    • Next.js
  4. 2023 — 2024Iloilo City

    Full-Stack Developer

    UniHub

    Our founding student team built a university marketplace — and got selected for DICT's STEP UP programme, one of 11 startups from 24 applicants in Western Visayas.

    • DICT STEP UP
    • Demo day pitch
  5. Origin · 2022 — 2026

    BS Software Engineering

    Central Philippine University

0314 projects

Selected Works

Ride a line — light up every project that uses it

 

  1. I

    Rare EarthRanger

    Real-time conservation intelligence at planetary scale

    Rare.org · Frontend Developer

    I built the frontend for Rare's EarthRanger Dashboard — a complex, data-driven analytics platform used by conservation teams worldwide. It takes real-time telemetry from GPS collars, camera traps, ranger patrols and remote sensors, and turns it into interactive maps, advanced analytics and actionable intelligence.

    Deployed across 750+ conservation sites, tracking 23,000+ animals globally

    • Next.js
    • TypeScript
    • Tableau Integration
    • Real-Time Data Streams
    • GIS Mapping
    • Advanced Analytics
    er-rare-dashboard.vercel.app↗
    Plate withheldThe dashboard lives behind Rare's login.
  2. II

    Mentoria

    Where AI becomes the teacher kids actually want

    Mentoria.gr (EU — Greece) · Full-Stack Engineer

    I designed and built an AI-powered education platform for one of Greece's leading education organizations. Students work with AI mentors through an agent-tool chat that adapts to their pace, and progress through structured learning modules in conversation.

    Serving students across the Greek education system

    • Next.js
    • TypeScript
    • AI/LLM Integration
    • Agent Architecture
    • Module System
    • Real-Time Chat
    project-mentoria.vercel.app↗
    Screenshot of Mentoria
  3. III

    OZ Tech

    Agentic workflows for everyone, from scratch

    OZ Tech (Current Employer) · Full-Stack Engineer — Agentic AI· Current

    I build agentic AI workflows for OZ Tech's clients — GoHighLevel (GHL) automation, custom-coded applications, and end-to-end AI agent pipelines. Every client gets a bespoke system, not a template.

    Multiple production systems serving active client base

    • Next.js
    • TypeScript
    • GoHighLevel
    • AI Agents
    • Custom APIs
    • Workflow Automation
    Screenshot of OZ Tech

Further volumes & short stories

  • Screenshot of RoboGuide
    Vol. IV

    RoboGuide

    IoT learning, demystified

    XzenTech Solutions · Lead Developer

    An IoT learning hub with a hardware wiki, guided tutorials, and community projects for Arduino, ESP32, sensors, and more.

  • Screenshot of XTS Web
    Vol. V

    XTS Web

    The digital workbench for engineers

    XzenTech Solutions · Lead Developer

    An engineering solutions and fabrication platform — e-commerce for hardware kits, custom fabrication services, expert consultation booking, and a community build-log feed.

  • Screenshot of StatSmart
    Vol. VI

    StatSmart

    From data to the right method, in seconds

    XzenTech Solutions · Lead Developer

    An AI-powered statistical analysis tool. Upload a dataset, and StatSmart identifies variables, recommends best-fit statistical tests, and walks you through the analysis.

  • Screenshot of Better Teams ROI
    Vol. VII

    Better Teams ROI

    How much is your team leaving on the table?

    An interactive ROI calculator for Better Teams — a quick-assessment tool that quantifies the cost of team dysfunction with personalized insights.

  • Screenshot of XZEN Connect
    Vol. VIII

    XZEN Connect

    Talent matching with algorithmic precision

    XzenTech Solutions · Lead Developer

    A professional talent marketplace with AI-powered matching, verified profiles, and secure escrow payments.

  • Screenshot of Proxima LMS
    Vol. IX

    Proxima LMS

    Learning management, built for robots

    XzenTech Solutions · Lead Developer

    A full-stack LMS designed specifically for robotics and technical education. Students submit code via an embedded Monaco editor, record video demos of physical robots, and check out hardware kits. Teachers review code and video side by side, grade with feedback, and track class-wide performance. Admins manage user accounts, hardware inventory, and tiered lesson packages. The curriculum structure supports real depth — every lesson has slide content, every code task has a working skeleton and rubric, every quiz has scored questions.

  • Plate withheldThe link is CPU's story on our STEP UP selection.
    Vol. X

    UniHub

    A student marketplace, built by students

    A university marketplace where students buy and sell, which I built with a team of CPU Engineering and Computer Science students. We were selected for DICT's STEP UP (Startup-tailored Empowerment and People-focused Upskilling Project) — one of 11 startups chosen in Western Visayas out of 24 applicants — and pitched at the program's demo day.

  • Plate withheldNot public yet.
    Vol. XI

    Cartman PH

    Deliveries, routed intelligently

    XzenTech Solutions · Lead Developer

    A delivery and logistics platform with live GIS mapping and optimized map-routing for deliveries — built across web and mobile.

    Private
  • Plate withheldPrivate tooling — it runs behind my admin login.
    Vol. XII

    Job Radar

    An agent that reads the whole job market

    A job-search agent that queries nine job boards in parallel, normalizes and de-duplicates everything into one feed, then scores each posting against a weighted skill profile — including whether the role is actually open to an applicant in my region, and whether the seniority matches. The same engine powers the live agent in Chapter III: type a role there and you are watching this run.

    Private
  • Screenshot of d8chaos
    Short story

    d8chaos

    Swarm intelligence meets dating chaos

    What happens when you throw a bunch of people into a dating pool and let AI simulate the chaos? Users create deep personality profiles — MBTI, enneagram, Big Five, attachment style, love languages, texting habits, dealbreakers, and a chaos factor — or just type a celebrity name and let the AI generate one. Once 3-8 profiles are in the pool, a multi-agent swarm simulation runs for 72 simulated hours producing matches, messages, ghostings, situationships, rage-deletes, and vague story posts. The primary engine is MiroFish, an open-source swarm intelligence platform that spawns autonomous agents with their own personality, memory, and decision-making logic — each interacting on simulated social platforms via a 5-stage pipeline: seed document upload → GraphRAG knowledge graph → agent persona generation → OASIS dual-platform simulation → ReportAgent analysis. Graceful degradation falls back to Gemini 2.5 Flash for faster single-call simulations. Results display as a color-coded timeline feed and narrative report. Shareable profile sets let friend groups compare what happened in their respective simulations — same profiles, different chaos.

  • Screenshot of Senti.AI
    Short story

    Senti.AI

    Emotional damage, quantified

    An absurdly over-engineered emotional damage assessment system that psychoanalyzes Filipinos through their music taste. Pick your songs, pick your MBTI, attachment style, love language, zodiac — the AI does the rest, and it does NOT hold back. Full-stack Next.js + TypeScript powered by Claude API with custom Taglish prompt engineering, Spotify OAuth (PKCE), AI-powered song classification, shareable IG/FB story cards, 1v1 match challenges, barkada group mode, a public leaderboard, and an admin analytics dashboard. 350+ song database covering OPM, K-pop, and international hits. Rate-limited so the API bill doesn't become its own emotional damage.

04Don't take my word for it

Exhibits

  • Exhibit A

    750+

    conservation sites running the EarthRanger dashboard I built the frontend for

    Source · Rare.org · 2025

  • Exhibit B

    6

    production web & mobile apps I've shipped as Lead Developer

    Source · XzenTech Solutions

  • Exhibit C

    0.985

    AUC-ROC of my thesis' CompGCN integrity layer at scoring which relationships are plausible

    Source · Undergraduate thesis

  • Exhibit D

    +423%

    faithfulness over a vector-RAG baseline, from the same framework

    Source · Undergraduate thesis

  • Exhibit E

    11 / 24

    startups selected in Western Visayas for DICT's STEP UP — my team's UniHub was one

    Source · DICT · 2023–24

  • Exhibit F

    6

    Anthropic Academy certifications I've earned — MCP, Agent SDK, Claude Code and more

    Source · Anthropic Academy

05My undergraduate thesis

Research

A GNN-Augmented Framework for Semantic Integrity Validation and Grounded Reasoning in Professional Knowledge Systems

Validate-then-Generate — a CompGCN integrity layer inserted between retrieval and synthesis, so the model can only reason over relationships it has verified.

  1. Stage 1

    PDF Ingestion

    Schema-guided extraction pulls documents into a knowledge graph. Entity and relationship types are predefined, constraining the model to domain-relevant concepts instead of inventing categories.

  2. Stage 2

    Entity & Relationship Extraction

    Zero-shot extraction identifies typed entities and relations — PROPOSES, CONTRADICTS, EVALUATES — at temperature 0, so the same input always yields the same graph.

  3. Stage 3

    Graph Storage & Provenance

    Every entity and relation is annotated with its source document, creating an unbroken chain from generated narrative back to triplet back to the original PDF.

  4. Stage 4

    Vector Embedding

    Solves the GNN cold-start problem: newly extracted entities have no embeddings, so DistilBERT generates 768-dimensional feature vectors to initialize the graph.

  5. Stage 5

    CompGCN Integrity Audit

    A 3-layer CompGCN with a DistMult decoder scores every relationship 0–1 for plausibility. Triplets below the threshold never reach the LLM — the validation gate the whole framework turns on.

0.985

AUC-ROC

0.958

MRR

0.984

Grounding

0.983

Faithfulness

+75% grounding, +423% faithfulness over vector RAG

  • Python
  • PyTorch Geometric
  • CompGCN
  • Neo4j
  • DistilBERT
  • Gemini 2.5 Flash
  • FastAPI
  • Next.js
06Coffee's on me

Coda

Building something that has to think? Let's talk.

CV · PDFor open your mail app →