Crafts [/ˈkræfts/] n - A record of my works, past and present, among which I count the shaping of this very site.
Of the Work I Do
I am a laborer in Information Technology, working with my hands as much as with my mind. I maintain, deploy, and contrive processes, monitoring services, and bespoke devices, all of which serve the efficiency of data centers and the effectiveness of those who keep them, the world over.
Data center power guardian [/ˈdata ˈsenter ˈpower ˈguardian/] n - I made this project to hunt power anomalies; it watches Eaton ePDU G3, APC Smart-UPS SRT, and Schneider Galaxy UPS systems. In finding over-provisioned PDUs we saved energy, for some sites could not return excess power to the local grid, their infrastructure yet unbuilt.
All values herein are simulations, to be replaced with real APIs and SNMP pollers that push logs to Grafana or Prometheus. It touches no live device as written; modify it before you trust it in production.
The truer labor lay in tending sensors and adjusting PDUs, so that all systems ran correctly. Some days demanded a climb of 40 feet to mend a sensor by hand; on others the system healed itself, kept within its configured guardrails.
Data center energy saver [/ˈdata ˈsenter ˈenergi ˈsaver/] n - This project weds Fuse FLIR A310 thermal video and DS18B20 contact sensors to a Raspberry Pi, cutting hotspot detection latency by 60% and reactive cooling costs by 18%.
I tested a like architecture at a data center. To govern consumption you must reckon the mass of the air, the cooling system’s capacity, and the air’s heat capacity, thus:
The case study used a simulated reading, watched for a set period, then raised the setpoint by 2 degrees for 10 minutes; the particulars vary between data centers and between the spaces within one. Large tier 3 sites require professional solutions, hundreds of sensors to govern and find anomalies in the cooling and electrical systems.
Much of the work was maintenance: sensors, Raspberry Pi boards, the repair of pipes and leaks in the cooling plant, all done amid daily checks and PMVs. Greater scale brings more sensors, more points of failure, more upkeep. Automation quickens deployment but lifts no responsibility from the operator; I grew more proficient as I took on more skills and more duties.
Intune Self-Heal [/ˈintune ˈself-hiːl/] n - This project finds devices with enrolment issues and repairs them of its own accord: device synchronization, MDM token refresh, enrolment token recreation, and the restarting of the Intune management extension, with all logs sent to Grafana afterward.
DC Power Monitor [/ˈdk ˈpower ˈmonitor/] n - This project is a lightweight system that watches data center power and temperature through existing SNMP-enabled PDUs and UPS devices, so that power costs and downtime both lessen.
It gathers power, voltage, current, and temperature via SNMP, keeps them in a local SQLite database, and serves a web dashboard of real-time data and historical charts; alerts sound when a value crosses its configured limit, and a mock collector mode permits trial without real hardware.
Its results are practical. It finds overloaded circuits before they disconnect, under-utilized PDUs fit for consolidation, temperature hotspots that waste cooling energy, and the trends by which UPS capacity is best shaped, all from one dashboard, without a clutter of vendor-specific tools.
DC Cooling Optimizer [/ˈdk ˈkuːlɪŋ ˈoptimizer/] n - This project is a lightweight advisor. It reads historical temperature and power data from CSV or SQLite, reckons statistics for racks and aisles, and offers concrete optimizations to decrease energy costs.
It finds hot spots, cold spots, and racks over- or under-utilized, then recommends: raise cold-aisle setpoints where temperatures run low, consolidate workloads into available capacity, balance airflow between aisles. Each counsel carries its estimated saving in kW and cooling cost.
Cooling can consume 30 to 50 percent of a data center’s energy, whence small setpoint adjustments yield large savings without breaching safety limits; and capacity found is equipment not yet bought.
DC Cable Audit [/ˈdk ˈkable ˈaudit/] n - This project is a command-line interface (CLI) tool that discerns data center cabling topology by LLDP and CDP, finds anomalies, and suggests repairs, so that cabling incidents and the cost of downtime diminish.
It connects to switches via SSH and reads neighbor tables into a structured topology, discovering missing or duplicate neighbors, speed mismatches between connected ports, and ports with no neighbor at all; findings and suggested repairs are exported to JSON.
Thus it spares the slow toil of tracing cables by hand, catches accidental loops before they breed broadcast storms, and brings undocumented connections and wasted port capacity into the light.
Network Switch Patcher [/ˈnetwork ˈswitkh ˈpatkher/] n - This project is a safe CLI tool that sends quick repairs for known issues on Juniper and Cisco switches. It keeps a library of vendor-specific remedies, connects via SSH, and delivers the correct commands, with a dry-run mode to examine changes before they are applied and a verification step to confirm the repair held.
The remedies answer common operational ills: recovering err-disabled ports, enabling BPDU guard on access ports against loops, enabling LLDP globally, wrapping Juniper configuration changes with commit confirmed, enabling jumbo frames. A mock mode permits rehearsal without live devices.
The aim is to end repeat outages quickly, without visiting each switch in turn; the patcher applies a tested, version-controlled repair and verifies it, shortening the time to repair and barring the errors of hasty typing.
NTP server [/ˈen-tiː-piː ˈsərvər/] n - This project raises a low-cost NTP server for the pool.ntp.org network. It keeps the chrony daemon within a Docker container and deploys that container to Kubernetes. Terraform builds the infrastructure, on AWS or on GCP, and Kustomize overlays hold the two clouds’ configurations apart. A network load balancer gives the server a public address, whereupon it answers the time requests that arrive on UDP port 123.
GitHub Actions pipelines test the code and carry it to deployment. They authenticate by OIDC and keep no stored cloud keys. The trials are shellcheck, hadolint, Terratest, and assertions written in plain bash; all must run before each deployment. The operators watch stratum, root delay, and time offset with the chronyc commands.
Neural style transfer [/ˈnyuːrəl ˈstīl ˈtransfər/] n - This project carries the artistic style of one image onto another. I made it in 2018 as a Bachelor of Engineering project: a Lua and Torch7 implementation of the Gatys algorithm, which draws its features from the Network in Network (NIN) model. NIN holds some 7.6 million parameters against VGG-19’s 143 million, and so the smaller model can work images up to 1080 pixels on a GPU with but 4 GB of memory.
The program coaxes the pixels of a generated image by gradient descent, lessening a content loss and a style loss together; the style loss reckons upon Gram matrices of feature correlation. Two Python tools preserve the colors of the content image: one transfers the luminance, the other the linear color distribution.
This website [/ˈthis ˈwebsite/] n - I shaped this website from the original Quartz 5 repository, and I thank its contributors and maintainers. Quartz 5 is a tool for deploying websites, letting users carry their Obsidian notes onto their own pages; through it I learned to write better software.
I trued the site into a dictionary: all notes moved into alphabetical folders (A to Z), a script affixing phonetic transcriptions to every term, a two-column layout for the lexicon pages, and a running head bearing the page number. You may travel this page by keyboard shortcuts: Shift+W for search, Alt+U to change the theme.
Further Works
- A Small Alarm for the UPS [/ˈraspberri ˈpi ˈalarm ˈfor ˈups ˈalerts/] n - A small physical alarm panel, built of a Raspberry Pi and SNMP UPS monitoring.
- Data center maintenance tasks [/ˈdata ˈsenter ˈmaintenanse ˈtasks/] n - A practical checklist of inspections and exercises by which a site is kept reliable.
Of Bespoke Devices and Efficiency
The works above I built from small, off-the-shelf parts. I have begun to set down the method in Bespoke Devices for the Datacenter, which tells how Raspberry Pi and Arduino devices do work that vendor tools will not. On the Making of Bespoke Devices follows the path from idea to deployed device; On Early Detection sets out the philosophy of monitoring; and Efficient datacenter case studies records how the largest operators pursue efficiency.