I engineered an edge-to-cloud sensor calibration stack using MQTT/Mosquitto, Telegraf, and InfluxDB, training RF/XGBoost regressors to map Alphasense O₃, NO, CO, NO₂, and SO₂ signals to reference-grade measurements. Managed experiment tracking and artifact versioning with MLflow and MinIO, instrumented the pipeline with Prometheus, Loki, and Grafana, and validated model robustness using SHAP, ALE, PDP/ICE, and DBSCAN-based outlier filtering against EU air-quality data requirements.

Building the AI document-intelligence layer for regulated, high-stakes back-office operations. I worked on the data intelligence side, dealing with idempotency and deduplication, canonicalization, observability, and lineage of important documents. I also architected the entire Brikli technical system end to end. As an engineer, I aim to make the ecosystem scalable in production across multiple cloud providers, keeping the stack lean and decoupled, with provider-agnostic abstraction boundaries.

I ran analysis and benchmarking of SAM 2 and SAM 3 on the VISOR dataset, generating per-frame IoU evaluations through Modal and Weights & Biases, architected inference optimizations using speculative decoding with draft models for temporal scanning and target VLMs such as LLaVA-NeXT and Qwen-VL, and evaluated a frozen InternVideo encoder with trainable MLP heads producing K+1 logits for per-frame state classification via softmax and argmax.

Automated production-grade cloud applications on Microsoft Azure using ARM-based Infrastructure as Code, with CI/CD deployment through Azure DevOps Pipelines and GitHub Actions. Managed enterprise identity, device, and application security using Microsoft Entra ID, Intune, MDM, MAM, and Privileged Identity Management, while developing administrative and provisioning automation with PowerShell, Azure CLI, Microsoft Graph API, n8n, and CIPP.