Selected builds, from production platforms to competition robots.
A tutoring platform in production use, a multi-GPU JAX implementation of STDE, a multimodal retrieval system, a rookie FRC robot whose team received the Rookie All-Star Award at the FIRST World Championship, and a transaction-monitoring model trained on 32M+ transactions.
01 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.
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).
Stack
JAX / CUDA / Multi-GPU (pjit/pmap)
03 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.
Team AlphaZ, Rookie All-Star Award Ankara Regional, FIRST Robotics Competition, 2026
04 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.
Stack
Java/WPILib / Computer Vision (YOLO) / Embedded Systems
05 Mar 2024
Anti-Money Laundering Detection
A production-orientedanomaly-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.
