Skip to content
View jjjardev's full-sized avatar

Block or report jjjardev

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
jjjardev/README.md

Connect with Me

GitHub Hugging Face LinkedIn Email


About Me

I'm a BSIT graduate from Negros Occidental, Philippines, building language technology for low-resource Philippine languages and deploying production-grade AI to edge devices with zero cloud dependency. I design systems for agents, not just for humans.

What I've built:

  • FiliSenti — Unified Tagalog-Hiligaynon sentiment model (0.8915 F1). Quantized to INT8 ONNX and deployed as a polished Flutter app. Runs entirely on-device.
  • HoardCore — Agent Toolkit for retrieval and deep research. Zero-API, zero-embedding, fully local. Runs [V] provenance-tagged research loops.
  • HiliSenti v1 — First multi-domain sentiment dataset for Hiligaynon. 23,337 sentences, 93.5% accuracy, permanent DOI.
  • TagaSenti — Multi-domain Tagalog sentiment dataset with adversarial augmentation. 35,686 sentences, 0.849 F1.
  • TalaKo — Offline-first credit tracking app for sari-sari stores. Flutter, SQLite, Material 3.
  • CPSU Server — Deployed a full production Linux server for my school: LAMP stack, DNS, LXQt desktop, student records integration.

My methodology: Full ML lifecycle — data curation → model training → quantization → mobile deployment → agent tooling — executed entirely on free resources.

Current focus: Building the default open-source infrastructure for Filipino NLP — datasets, edge models, benchmarks, and agent tooling that compound into citable, grantable assets.


Tech Stack

Languages & Frameworks

Python Dart SQL Bash HTML5 CSS3

Mobile & Deployment

Flutter ONNX Android SQLite

Machine Learning & Data

PyTorch Transformers scikit-learn Pandas

Agent Tooling & Retrieval

SQLite FTS5 Ollama

Systems & Infrastructure

Linux Ubuntu Server Apache MySQL Docker Git

AI & Edge

Colab


Featured Projects

FiliSenti — Unified Filipino Sentiment Model & Mobile App

HF Model GitHub

  • Unified Tagalog-Hiligaynon sentiment model achieving 0.8915 macro F1 — the highest published score for 3-class Filipino sentiment.
  • 355M-parameter XLM-RoBERTa-large quantized to INT8 ONNX (537 MB) and deployed in a Flutter Android app with on-device inference.
  • Ported XLM-RoBERTa's SentencePiece tokenizer to pure Dart — solved a critical deployment barrier for mobile NLP.
  • Polished Material 3 app with text input, file upload, color-coded sentiment highlighting, and CSV/JSON export.
  • Full ML lifecycle demonstrated: dataset curation → model training → quantization → mobile deployment.

HoardCore — Agent Toolkit for Retrieval & Deep Research

GitHub

  • Zero-API, zero-embedding, fully local retrieval engine for AI agents. Built in 12 hours.
  • Hybrid retrieval: FTS5 (BM25-style) + lexical hashing (FNV-1a) → RRF fusion.
  • Provenance-tagged output: [V] (verified), [E] (external), [H] (hypothesis) — every claim is traceable.
  • DISCOVER → INGEST → RECALL → EMIT research loop. Runs deep research briefs at cheaply.
  • Single-file Python, zero dependencies, designed to be driven by AI agents.

HiliSenti v1 — First Sentiment Dataset for Hiligaynon

HF Dataset DOI

  • First publicly available multi-domain sentiment dataset for Hiligaynon (10M+ speakers).
  • 23,337 sentences across news, social media, cross-lingual translations, and synthetic edge cases.
  • XLM-RoBERTa-large fine-tuned: 93.5% accuracy, 93.4% macro F1.
  • Built on free Google Colab (Tesla T4) and free 15GB Google Drive.
  • Within 48 hours of release: 40+ downloads, 14,600+ combined Reddit views, 141 upvotes (97%+ ratio), unsolicited commercial interest.
  • arXiv paper in preparation.

TagaSenti — Multi-Domain Tagalog Sentiment Model

HF Dataset GitHub

  • 35,686 -sentence multi-domain Tagalog sentiment dataset (e-commerce, news, social media).
  • Targeted adversarial augmentation strategy: 3,766 rows across 8 failure categories based on a 100-sentence adversarial benchmark.
  • XLM-RoBERTa-large fine-tuned: 0.849 test F1, 92% benchmark accuracy.
  • Zero-shot transfer to Hiligaynon at 0.591 F1 — confirmed cross-lingual transfer within the Philippine language family.
  • Entire pipeline executed in 6 days on free Google Colab (Tesla T4) and free 15GB Google Drive.

TalaKo — Offline-First Credit Tracking for Sari-Sari Stores

GitHub APK

  • Production-grade offline-first Android app for sari-sari store owners to track customer credits and price references.
  • Complete SQLite database schema with customers, items, and credits tables.
  • Features: customer management, credit tracking with partial payments, overdue detection (14-day), searchable price catalog, backup/restore (JSON + raw DB).
  • Built with Flutter and Material 3 UI; ~1,200 lines of Dart code.
  • Supports adjacent use cases: tricycle operators, rice retailers, water refilling stations, 5-6 lenders, paluwagan organizers.
  • Motivated by personal experience managing a family sari-sari store with inconsistent notebook tracking.

CPSU Linux Server Infrastructure

2025 | Ubuntu Server · Apache · PHP · MySQL · DNS · LXQt

  • Designed and deployed a full production Linux server for Central Philippines State University — Hinigaran Campus.
  • LAMP stack (Apache, PHP, MySQL), custom internal DNS, and LXQt desktop environment selected for low resource usage and GUI familiarity for non-technical staff.
  • Configured remote FTP access enabling 5+ teaching staff to update web content without physical server visits.
  • Coordinated deployment and environment compatibility for a third-party Transcript of Records system integrated into the admin dashboard.
  • Awarded Best in On-the-Job Training for this work.

Research Impact

Project Type Scale Key Metric Status
FiliSenti Model + App 56,781 sentences 0.8915 macro F1 Public (HF + GitHub)
HoardCore Research Toolkit 275 docs, 5,086 chunks $0.0074 per deep research brief Public (GitHub)
HiliSenti Dataset 23,337 sentences 93.5% accuracy Public (HF, DOI)
TagaSenti Dataset 35,686 sentences 0.849 F1 Public (HF)
TalaKo Mobile App 49 MB APK ~1,200 LOC Public (GitHub)

All projects were built on zero budget — free Google Colab, free Google Drive, open-source models, and free LLM web UIs.


Let's Connect


Pinned Loading

  1. HoardCore HoardCore Public

    Give your agent a memory it can prove. One Python file: local SQLite vault, hybrid FTS5 + vector retrieval, Cloudflare-aware fetching, and a research loop whose [V]/[E]/[H] citations are machine-ve…

    Python 7 2

  2. filisenti filisenti Public

    On-device sentiment analysis for Tagalog and Hiligaynon. Achieves 89.1% macro F1. Built with XLM-RoBERTa-large, INT8 ONNX, Flutter. Full ML lifecycle

    Dart

  3. talako talako Public

    TalaKo: Sari-Sari Store Price Reference & Credit Manager. Fully offline.

    Dart