Data & AI Engineer · RAG Architect · GNN Researcher · Desktop Software (Tauri/Rust)
I am a Data & AI Engineer focused on building scalable data pipelines, automating complex processes, and deploying production-grade AI systems. I also build native desktop apps with Tauri + Rust. Currently an Intern at Izii Soluções Tecnológicas, Tech Lead for Project Metis, and founder of Mora Org — an open-source collective building software with intention.
The best place to see my full trajectory and live demos: 👉 cesarsibila.vercel.app
| Category | Tools & Technologies |
|---|---|
| Languages |      |
| AI & ML |          |
| Data & Infra |         |
| Desktop |   |
| Automation |   |
- ⚡ Metis: A full-stack LoL analytics platform (p-0.9.22). Features champion pages with build/matchup/synergy analytics, z-score tier ranking, match scoreboard with interactive timeline charts, a custom design system with PT/EN i18n, and a Medallion data pipeline (Riot API → Cloudflare R2 → Supabase). AI-powered tactical coaching via Gemini 2.5 Flash with RAG guardrails; an agentic hybrid-search pipeline (KNN + BM25/RRF) via OpenRAG (Langflow + OpenSearch + Docling) is in active development. Stack: FastAPI, Next.js, Polars, pgvector, GitHub Actions.
- 🕵️ GNN Fake News: Academic research project (TCC — PUCSP). Fake news detection on Bluesky via network topology, comparing GCN, GAT, and GraphSAGE architectures against textual baselines. Trained on UPFD FakeNewsNet + GossipCop datasets and applied to real AT Protocol firehose data. Includes ablation studies, cross-dataset benchmarks, GNNExplainer interpretability analysis, multilingual topology tests (PT/EN/DE), and a full interface (Next.js + FastAPI + Supabase). Key finding: topology alone reaches F1=0.81 on GossipCop; dual-model rule (textual + topological) hits F1=0.97. Stack: PyTorch Geometric, sentence-transformers, Docker.
- 📚 cdia-m7-PUCSP: PUCSP CDIA M7 coursework. MLOps pipeline with synthetic data generators for credit/churn/fraud domains — scikit-learn models trained and published to HuggingFace via GitHub Actions. Also includes a production-grade FastAPI restaurant service with Pydantic v2 and routers.
- 🗄️ Atlas: Open-source headless CMS for modern admin panels, born from a Scientific Initiation (IC) project. Build typed tables visually — each schema becomes a real physical DB table. Features multi-tenant isolation, 3-tier access (master/admin/moderator), JWT + QR Auth, 6 themes × 4 modes, SQL/CSV/XLSX import, and public API toggle. Stack: Next.js, FastAPI, SQLAlchemy, Neon Postgres. Part of Mora Org.
- ✍️ Glyph: Open-source desktop software for visual authoring and advanced typography (formerly PEG). Currently at Phase 5/9 — export pipeline in development; Phases 0–4 (canvas engine, lettering, GIF, store) are done. Features per-character control, organic text effects (tremor, neon flicker), video-as-fill clipping masks, active pause blocks (VHS noise, black screen), and GPU-accelerated export (NVENC/AMF). Two tracks: Core (stable, lightweight) and Obsidian (experimental AI video inpainting via VOID Engine, 12GB+ VRAM). Stack: Tauri (Rust), Next.js, Fabric.js, FFmpeg. Part of Mora Org.
-
📋 LattesDirector: Native desktop app (Tauri v2) that connects your Google Calendar to your Lattes academic CV via a local AI agent (OpenClaw via WebSocket). Monitors diffs between real calendar events and Lattes entries, and presents a side-by-side approval dashboard — no manual bureaucracy. Features glassmorphism UI with Framer Motion, typed OpenClaw contracts (
CalendarEvent,LattesSuggestion,DiffPayload), isolated Gov.br auth via InAppBrowser, and a security red-teaming suite (Playwright) that validates auth tokens never leak towindow. Stack: Rust, React 19, Zustand, Vite, Playwright. Part of Mora Org. -
🎾 Heimdall: End-to-end tennis analysis pipeline that democratizes Hawk-Eye-style stats using a single phone or YouTube video. Uses TrackNetV2 (U-Net CNN that processes 3 consecutive frames to track fast, motion-blurred objects — far superior to YOLO for this domain), homography matrices to convert 2D pixel coordinates to real 3D court meters, and a physics engine that detects hits/bounces via velocity derivatives and classifies topspin vs. slice via the Magnus effect (
$g > 9.8, m/s^2$ = topspin,$g < 9.8$ = slice). Stack: Python, OpenCV, PyTorch, Pandas, Savitzky-Golay/Kalman filters. - 🗺️ Travel-Salesman-Problem: A collection of heuristic and exact algorithms for the classic TSP, including implementations of greedy nearest-neighbor, 2-opt local search, and dynamic programming approaches.
- 🃏 Ze_Pelintra: A study on Parallel Computing and Monte Carlo Tree Search (MCTS) applied to the Brazilian card game "Truco" — exploring how parallelism affects game-tree search depth and decision quality.
- LinkedIn: linkedin.com/in/cesarsibila
- HuggingFace: huggingface.co/Zaras210
- Mora Org: github.com/Mora-Org
- Email: cesarsibila210@gmail.com
"Turning raw data into strategic intelligence."