ArcSync — Architecture-Aware AI Specification Agent Full-stack multi-agent system generating a technical spec from any codebase in under 30s. 2nd of 300+ teams, IBM BOB Hackathon 2026

• Architecture: 4-agent orchestration (IBM Watsonx.ai · Granite 3 8B) with a hybrid RAG pipeline — architectural file weighting, complexity scoring, domain synonym expansion
• Key Features: FastAPI backend · 50MB ZIP ingestion · JSONL audit logging · 13+ languages supported

LARF — Legal Adaptive Routing Framework 5-stage LLM pipeline answering OFW legal questions across Philippine and Hong Kong law. Delivered to the Department of Migrant Workers under a formal MOA with Saint Louis University.

• Architecture: Normalization → routing → hybrid retrieval (FAISS + BM25 + RRF) → generation → safety-auditing
• Key Features: Coarse-to-strict reranking stage · low-confidence/unsafe-output flagging for human review

Incident Response Copilot — DevOps/SRE Multi-Agent RAG System 🚧 In Progress · Week 1 of 6

• Architecture: Fine-tuned cross-encoder reranker (MLflow-tracked) over VOID postmortems + runbooks, orchestrated with LangGraph
• Key Features: Triage → RCA → remediation → guardrail agents · Dockerized · GitHub Actions CI/CD · deploying to AWS

Credit Risk Modeling — Lending Decision Support System Calibrated LightGBM (Platt scaling) on 307k thin-file applicants, paired with a SHAP narrative engine explaining declines in plain English.

Telecom Churn Prediction — End-to-End ML Case Study LightGBM + CatBoost ensemble, 100k records, 140 features. Cost-sensitive threshold tuning moved recall from 65.7% → 91.5% — $5.18M net annual benefit.

Fraud Detection — Cost-Sensitive Transaction Modeling 🔄 Ongoing RandomForest on 6.3M imbalanced PaySim transactions, scored against a custom Net Error Revenue metric built on Philippine fintech cost assumptions.