Marshall Rasendria Mahendra

Real-time sign language recognition and 3D avatar platform for BISINDO

• All camera inference runs on-device, so no video ever leaves the browser.
• 26-class alphabet classifier at 99.31 percent validation accuracy, plus a 32-class word-level 1D CNN at 83.13 percent signer-independent accuracy running at 15 fps with 6 ms median latency.
• Rigged 3D avatar performs sign poses derived from dataset landmarks via analytic inverse kinematics.

Next.js /TypeScript /TensorFlow.js /MediaPipe /react-three-fiber

Deepfake detection and digital media forensics system to stop AI-powered fraud

• Trained a custom deepfake detection model from scratch on 10,000+ images and videos, reaching 90+ percent detection accuracy.
• Built the entire web application independently, end to end.
• Delivered as a social impact initiative educating 40+ residents in a rural village about AI-generated media fraud.

PyTorch /OpenCV /Next.js

RAG-powered API intelligence and gap analysis platform

• Scans a codebase from a GitHub URL or dropped folder, detects backend routes across Express, FastAPI, Laravel, NestJS, and Next.js and frontend calls across fetch, axios, and React Query, then flags connected, orphan, and ghost endpoints.
• “Ask your codebase” RAG: route-handler chunks, overlapping 60-line windows, and per-route summary cards embedded with Gemini into pgvector behind an HNSW cosine index, retrieving the top 6 after similarity thresholding and overlap deduplication.
• Answers are grounded strictly in the retrieved sources with inline citations to file paths, line ranges, and routes, and say so plainly instead of inventing endpoints when nothing relevant comes back.
• Hybrid pipeline runs deterministic parsers first with LLM fallback, ships with Docker on Kubernetes, and is backed by 200+ Vitest tests.

Next.js /TypeScript /Groq (Llama 3.1) /Gemini embeddings /pgvector /PostgreSQL /Docker /Kubernetes

Multi-domain plugin that turns AI coding agents into a senior ML research engineer

• Auto-routes a plain-language task to the best method across computer vision, medical imaging, NLP, tabular, time-series, and more.
• Encodes a research-first, leakage-aware workflow grounded in 2021 to 2026 literature, with realistic-ceiling estimation and split audits.
• Ships an iterative accuracy-improvement loop with per-problem metric and explainability selection.

Claude Code /Antigravity /23 skills /14 domains

Controlled Comparison of Multimodal Hybrid Deep Learning for Benign-Malignant Mammogram Classification on CBIS-DDSM

JADIMI: Journal of AI-Driven Informatics and Management Information

Vol. 1 No. 1, August 2026, pp 20-33 · First author

Marshall Rasendria Mahendra, Farhan Muamar Fawwaz, Naisya Aghis Nabila, Muhammad Rakha

Many deep learning studies on CBIS-DDSM report accuracy above 95 percent, yet they partition the archive at image level, so views of one patient land in both the training and the test set and the figures cannot be trusted. This paper rebuilds the comparison...more