ARA: RAG Chatbot for Abrasea
Built end-to-end RAG pipeline for Abrasea.com, a
government-funded mangrove conservation platform.
Designed a dual-source knowledge base combining platform-specific documentation and peer-reviewed scientific
journals on mangrove ecology and blue carbon,
embedded with BGE-M3 (1024-dim) and stored in Pinecone.
Orchestrated the full agentic workflow on n8n with
Nvidia Nemotron Ultra 3 as the inference model. Evaluated across 100
domain-specific queries using a custom RAGAS framework, validated by two frontier judges (GPT-5.5
and Claude Sonnet 5),
achieving Faithfulness 0.99, Answer Relevancy 0.99, and
Context Utilization 0.91 at $0.003/query.
RAG
n8n
Vector Database
LLM-as-a-Judge
Agentic AI
View Live →
Conjify: Anemia Screening via Conjunctival Imaging
Built the full AI pipeline for Conjify, a mobile-first web app for non-invasive anemia
self-screening via smartphone camera. I designed and trained a dual-head
EfficientNet-B0 that simultaneously classifies anemia presence and estimates hemoglobin levels
quantitatively, trained on CP-AnemiC dataset (710 images) using multi-task learning
with progressive unfreezing. Integrated Grad-CAM to extract spatial activation
statistics for structured Llama 3.3 70B clinical interpretation in plain Indonesian.
Model achieved 84% accuracy, AUC-ROC 0.902, and Hb regression MAE 1.515 g/dL. Deployed on Hugging Face Spaces via Gradio API.
Computer VisionEfficientNet-B0Multi-task
LearningGrad-CAMLLMGradio
View Repository →
Unspun: Bias-Aware Product Search Engine
Capstone project built during my internship at Flyrank AI, also submitted to the
DevNetwork [API + Cloud + AI] Hackathon 2026 under SerpApi "Best AI Use Case" track. Designed a parallel retrieval pipeline that concurrently fires organic search, Reddit-scoped
sentiment, and Google Shopping pricing via SerpAPI, feeding a deterministic bias audit layer that quarantines affiliate-heavy and listicle domains before any LLM
sees the results. A single Cerebras gpt-oss-120b inference call then handles the
full synthesis: sentiment ranking, astroturf flagging, savings delta computation, and quarantine reason
tightening. Post-render, a separate Google Trends endpoint resolves interest curves
per ranked product. Engineered with hard latency budgets (6.5s parallel cap, 9s synthesis ceiling) and
graceful degradation so Reddit and Shopping failures degrade to empty rather than blocking the organic
pipeline. Deployed on Vercel with Next.js App Router +
TypeScript frontend and FastAPI Python serverless backend, zero database.
TypeScript
FastAPI
Cerebras
SerpAPI
View
Repository →
NutriVision
Built as the AI component of a cross-path capstone at Coding Camp DBS x Dicoding. I
implemented EfficientNetV2 as the backbone with BiFPN for multi-scale feature fusion and FCOS as the
detection head, a fully anchor-free approach using stride-based convolution instead of predefined anchor
boxes. The app identifies fast food items from major brands and estimates their exact nutritional content.
Computer VisionEfficientNetV2FCOSBiFPNAnchor-Free Detection
View App
→
NALAR: Media Literacy Simulation
Built an agentic AI simulation platform where specialized LLM
workflows autonomously generate scenario briefings, audience personas, editorial assets, network
propagation, and post-incident analysis. Designed a multi-stage AI orchestration
pipeline with structured JSON outputs, prompt specialization, and multi-model
routing across Gemini 2.5 Flash and Groq, monitored through Langfuse
observability. Engineered a real-time AI-driven propagation engine that
recursively simulates branching social interactions on a force-directed graph, alongside token-budget
guardrails, hybrid rule-based + LLM evaluation, and secure server-side orchestration using Next.js App Router and Supabase.
Next.jsTypeScriptAIGroqGeminiSupabasereact-force-graph-2d
View App →
Personal AI Agent (WhatsApp)
Built and deployed a personal AI agent directly into my own WhatsApp, handling automated responses, task assistance, and
context-aware interactions through a bot pipeline running on personal
infrastructure.
AI AgentWhatsAppAutomationPython
PyGrind
A browser-only, 100% AI-driven coding practice platform for AI/ML and backend
engineers. Every task, curriculum path, and code review is generated dynamically by an LLM (Groq or Gemini) using the user's own API key, with no backend
and no database. Built an adaptive tier system where problem difficulty scales per topic, a Monaco-based practice loop with real-time AI code review gating progression, and an
owner-maintained HTML handbook synced as reference material for every problem.
ReactTypeScriptMonaco EditorLLM IntegrationZustandTailwind CSS
View App →
LexiLSTM: Campus Complaint Classifier
The real challenge was not the model, it was the absence of labeled data. I used weak
supervision with a domain lexicon and fuzzy string matching to auto-generate pseudo-labels across six
complaint categories from informal Indonesian social media text. I compared standard BiLSTM, BiLSTM +
self-attention, and BiLSTM + multi-head attention. The key finding: added complexity does not consistently
win when training labels are noisy. The hybrid approach still delivered +18%
accuracy over baseline, with attention weights providing interpretable token-level insight.
NLPBiLSTMSelf-AttentionWeak SupervisionPython
View Deck
→
Hydros: AI-Powered Urban Water Investigation Platform
Built for the IEEE OneAquaHealth Global Hackathon 2026, an evidence-first
investigation tool that turns a photo and location of an urban waterway into a structured risk assessment
grounded in the One Health model. Enforced a strict observation
→ evidence → inference separation in the type system itself, so an AI-flagged observation
can never silently become a contamination claim — only a dedicated risk-assessment layer may conclude,
and “insufficient data” is a first-class outcome. Designed a multi-stage pipeline: NVIDIA NIM for descriptive visual analysis, human-in-the-loop confirmation of every
observation, Cerebras gpt-oss-120b for research planning and web-evidence synthesis
plus One Health exposure-pathway mapping (each pathway cited and conditional), and Groq for the final reasoning step. Added deterministic degradation alerts, geohash-based
site clustering with risk trend tracking, and FHIR R4 / JSON-LD
(FAIR-compliant) data exports. Deployed with Next.js, TypeScript, Supabase, and MapLibre GL.
Next.jsTypeScriptCerebrasNVIDIA NIMGroqSupabaseMapLibre GL
View App →