// projects

Async video intelligence pipeline processing media streams through modular detection workers (face recognition, object detection, OCR) using Celery + Redis. PostgreSQL for metadata indexing, FAISS for vector similarity search, REST API for downstream consumers. Deployed on AWS EKS with GitLab CI/CD.

Production-grade home network running OPNsense firewall with Tailscale overlay, Pi-hole + Unbound as a full recursive DNS resolver, Caddy reverse proxy, and multiple self-hosted services. Wardriving setup with Kismet on Kali/Raspberry Pi. NixOS/Nix for reproducible dev environments across machines.

SMS-based field survey reporting application built with FastAPI, Twilio, and MongoDB. Used in real data collection workflows for structured reporting from the field.

Distributed pipeline for OCRing millions of PDFs on self-hosted hardware instead of paying for cloud compute. Docker Swarm workers pull jobs from RabbitMQ and write extracted text and metadata to MongoDB for downstream cleaning. A follow-on Kubernetes stage runs Celery workers on Valkey, autoscaled by queue depth with KEDA, to generate text embeddings with a locally served model via Ollama.

Company infrastructure as code: a Talos Linux Kubernetes cluster with Helm-managed monitoring, alerting, Spark batch jobs, and controller-driven upgrades; a self-hosted NetBird control plane (custom Helm chart) replacing Tailscale; a versioned Python CLI for UniFi firewall policy that is dry-run by default, snapshots before changes, and verifies rule ordering; and reusable Forgejo Actions workflows for a Kubernetes runner with a Podman sidecar.

Distributed NLP pipeline that turns hyper-local news articles into structured event data. Documents are queued in Redis and fanned out to Celery workers, which use a locally hosted LLM through Ollama to extract records into Pydantic models. Results accumulate in a growing dataset that is batch-published to a geospatial data platform via its API. Containerized with Docker Compose and Helm for Kubernetes.