2025 – Present
AlgoTrack
A tutoring and education platform for tracking student progress, attendance, and pbInfo problem-set performance, built end to end and deployed in production.
- Full-stack platform (Next.js, PostgreSQL, Railway) covering student progress tracking, attendance, tutoring session management, and cohort analytics.
- Companion Chrome extension integrates with pbInfo to pull problem-set and submission data automatically.
- Deployed and actively used in production at algotrack.deonise.ro.
Next.jsPostgreSQLRailwayChrome Extension
May 2025
Stochastic Taylor Derivative Estimator for High-Order Differential Operators 2
Multi-GPU JAX implementation enabling scalable high-order automatic differentiation, validated on high-dimensional many-body Schrödinger PDEs.
- Implemented a multi-GPU version of STDE in JAX using Taylor-mode AD, randomized jet sampling, and pjit/pmap primitives for cross-GPU parallelism.
- Achieved a 6.86× runtime speedup and 27% lower GPU memory footprint across 4×V100 GPUs by parallelizing jet-level contractions and minimizing inter-device synchronization.
- Sustained numerical precision to machine epsilon with near-linear strong scaling; results published and awarded Best Scientific Paper and Most Impactful Paper (Engineering).
JAXCUDAMulti-GPU (pjit/pmap)
Aug 2025
AtlasRAG · Multimodal Retrieval-Augmented Reasoning System
A multimodal RAG framework fusing LLaMA 3.2-70B and Mistral-v0.3 with knowledge-graph grounding and hybrid retrieval, built to cut hallucination and latency.
- Reached 94.1% factual consistency and a 45% reduction in hallucination rate vs. a GPT-5 baseline, combining LLaMA 3.2-70B and Mistral-v0.3 with Neo4j knowledge-graph grounding and hybrid FAISS/BM25 retrieval.
- Boosted cross-modal retrieval precision to 88% with text-image-audio fusion (CLIP-ViT, Whisper embeddings) unified through a LangGraph agentic planner.
- Lowered end-to-end latency by 61% (p95 = 1.3s) via LoRA fine-tuning, 4-bit GGUF quantization, and asynchronous vector-store caching in a containerized inference mesh.
LLaMA 3.2-70BMistralNeo4jFAISSLangGraphCLIPWhisper
2026
AlphaZ · FIRST Robotics Competition, Team #11141
Lead mentor for a rookie FIRST Robotics Competition team, from turret targeting software to a World Championship award, the team's first for a Romanian FRC squad.
- Developed a ballistic targeting system for turret control using projectile-motion equations and dynamic distance estimation, achieving <5% targeting error.
- Implemented autonomous detection, alignment, and firing routines, removing manual intervention during match operation.
- Trained and deployed a YOLO-based computer vision model for real-time object detection at 90% precision.
- Won Rookie All-Star (Ankara Regional), Quality Award (Başkent Regional), and Rookie All-Star at the 2026 FIRST Championship (Daly Division, Houston), the first Romanian team to win an award at a FIRST World Championship event.
Java/WPILibComputer Vision (YOLO)Embedded Systems
Mar 2024
Anti Money Laundering Detection
A production-oriented anomaly-detection system for transaction monitoring, trained on 32M+ transactions.
- Achieved 95% precision and 96% recall on 32M+ transactions with a 5-layer sparse autoencoder (BatchNorm, Dropout, MSE + L1 sparsity, AUCPR-optimized thresholding).
- Reduced false positives by 18% and improved convergence by 42% via timestamp normalization and cyclical learning-rate scheduling (OneCycle/Adam).
- Cut inference latency from 3.2s to 0.8s per transaction, deploying TensorFlow Serving behind Kubernetes HPA with Kafka streaming ingress.
TensorFlowKubernetesKafkaFeast
