Built by Tech Guard
Products we designed & run ourselves, & open source tools we publish for others to use.
NK NODE
Our agentic network security appliance, designed & built in house. NK NODE puts AI agents to work on network security from a device on your own network, engineered with the same security first foundation as everything we ship.
Open Source
Tools we publish for anyone to use, power to the people.
One Command Install
GLM-5.3-Flash on Two DGX Sparks
One command stands up GLM-5.3-Flash as a fast, keyed, OpenAI compatible endpoint for your developers & agents. It is a recipe for teams that have the hardware, built on open work credited in the repository.
Measured on our own pair of Sparks: 90 to 109 tokens per second on JSON & structured output, & 80 to 86 tokens per second combined across four requests at once.
curl -fsSL https://raw.githubusercontent.com/TechGuardCoders/glm53-tensorfold-spark-recipe/main/install.sh | bash -s -- --worker <user>@<worker CX7 IP>Add --dry-run to check both nodes before anything changes.
What You Need
- Two NVIDIA DGX Sparks, cabled CX7 to CX7
- Docker on both nodes
- Passwordless SSH from the head node to the worker
- About 400 GB free on each node
- One to two hours for the first build & a 176 GB weights download on each node
License
Apache 2.0. The optional DFlash2 drafter is licensed for non commercial use. Every credit & license is listed in the repository.
Repos For Developers:
TypeScript · MIT
Pocket Watching
Cost per token dashboard for a self hosted inference cluster: latency, GPU utilization, & dollars per million tokens by model, tenant, & route.
View on GitHub→Python · MIT
Batcher
Continuous batching load test that finds where throughput stops scaling & latency starts to climb.
View on GitHub→Python · MIT
Spinal
Request tracing through queue, prefill, & decode, with drift alerts when serving slips from its baseline.
View on GitHub→Python · MIT
Ball Knowledge
Multi model AI gateway: routing, retries, fallbacks, rate limits, & token budgets for each tenant.
View on GitHub→Python · MIT
Bastion
Secure multi tenant inference layer: hashed keys, scopes, a tamper evident audit chain, & sandboxed execution.
View on GitHub→Python · MIT
Furnace
Checkpointed distributed training with fault injection, proving a killed job resumes from its last checkpoint.
View on GitHub→Python · MIT
Green Zone
Speculative decoding teardown that measures draft acceptance & cost per accepted token.
View on GitHub→Python · MIT
Popcorn
One hot path rebuilt in C & measured against the original, behind a byte identical correctness gate.
View on GitHub→Python · MIT
Flux Capacitor
Queue based autoscaler with cold start mitigation & graceful scale down.
View on GitHub→Python · MIT
Circuiter
GPU partitioning lab: slices, weighted fair scheduling, hard limits, & a contention drill.
View on GitHub→Python · MIT
UberCode
Model weight delivery through a content addressed shard registry with resumable bring up.
View on GitHub→Python · MIT
Flunk
Multi region failover drill that measures recovery time & requests lost.
View on GitHub→Python · MIT
Blueprint
The capstone: every measurement from the repositories above, gathered with its source.
View on GitHub→Python · MIT
Archive Blogger
Multi agent blog writer where a planner & a writer each answer to a validation check.
View on GitHub→Every public repository lives at github.com/TechGuardCoders.