Projects
Three open-source tools, and the client infrastructure work behind them.
01 Open source
Three tools, published and installable. One argument runs through all of them:
the failure that costs you is the one nothing reported. A corpus that
silently lost documents. A Terraform repository that passes validate and
cannot apply. A GPU dashboard reading 100 percent while the machine
does almost nothing. In each case the system knew and did not say.
truthscale
nvidia-smi reports “GPU utilization” and NVIDIA defines it as the
percent of time in which at least one kernel was executing. It is a time-based occupancy
flag, not a measure of work. One small kernel on 1 of an H100’s 132 streaming
multiprocessors, running the whole sample window, reads 100 percent while the other
131 sit idle.
Autoregressive LLM decode is the pathological case: one small kernel per token keeps the flag pinned while SM occupancy sits near 15 percent and the tensor cores are close to idle. Production clusters average around 5 percent real GPU utilization behind dashboards that read healthy, and autoscalers then act on the number that caused the confusion.
truthscale reports the honest number beside the misleading one, expresses delivered work as a fraction of a measured ceiling rather than a datasheet figure, and explains in words why the two differ. A counter the driver did not return is reported absent, never as zero, because an idle GPU and an unsupported driver must not look identical.
Every command runs against recorded traces, so it can be evaluated on a laptop with no GPU attached.
websieve
Turns a web crawl into a dataset you would actually train on: boilerplate extraction, Unicode normalisation, the Gopher and C4 quality heuristics, exact and MinHash near-duplicate removal, and sharded output with verifiable manifests.
Two properties matter more than the pipeline. Every dropped document is attributed to the rule that dropped it, and the counts reconcile, so a filter you cannot explain is a filter you cannot tune. And the quality rules adapt to the writing system: the published heuristics were all derived from English and assume words are separated by spaces, so applied unchanged they reject ordinary Chinese, Japanese, Thai and Korean outright.
stackmason
Answer some questions, get a Terraform repository rather than a snippet: one state file per environment, providers pinned and configured with an explicit region, and every resource tagged.
The defaults are the product. A generator multiplies whatever it emits, so one insecure
default becomes a hundred insecure deployments run by people who reasonably assumed the
generator knew better. It therefore refuses to emit a data port open to
0.0.0.0/0, a publicly accessible database, a bastion with no source
restriction, or skip_final_snapshot in production, and there will never be a
flag that downgrades a refusal to a warning.
02 Client engagements
Contract and consulting work, delivered through Eprecisio Technologies. Client code is not mine to publish, so what follows is the work and the outcome rather than a repository link.
Stack8s, Kubernetes automation platform for AI/ML teams
- Cluster provisioning under ten minutes, production-ready with GPU compute across AWS, Azure and on-prem, with Kubeflow plug-and-play so a team can use the compute immediately rather than after a week of setup.
- Multi-tenant Kubernetes on Kamaji, with Gateway API for routing and traffic management across tenant workloads.
- Rancher-based cluster management across multi-cloud and on-prem, with GPU orchestration and workload distribution.
- Vanilla Kubernetes on customer-owned hardware, for data sovereignty in regulated AI and ML environments.
- Full OpenStack on bare metal with Ceph as the storage backend, via custom Ansible roles and playbooks rather than a containerised control plane: Nova, Neutron, Cinder, Keystone, Glance, Horizon and Ceilometer, with OVS, VXLAN and DVR networking, re-runnable for redeployment and cluster expansion.
Hyve Labs, real-time gaming platform
- 60 percent or more reduction in AWS spend, through model selection and inference-harness work on Bedrock, disciplined rightsizing of compute and storage, and operational practice across the footprint.
- Rebuilt the AWS data and ML stack (Bedrock, Airflow, Redshift, Glue) around a Kafka pub/sub architecture; event fan-out feeds analytics, model workflows and downstream game systems at real-time scale.
- Shipped two agentic harnesses in production: a Marvin-based one for game content generation, and a LangGraph SRE harness for incident triage, dynamic alert classification and token-optimised remediation.
- Owned the production security layer (layered IAM and RBAC, secrets management, image hardening, network policy), ran regression, load, stress and periodic penetration testing, and held PagerDuty escalation with P0 to P2 SLA enforcement.
- Owned end-to-end production: real-time event fan-out, elastic multi-region autoscaling, low-latency multiplayer backends and the release lifecycle.
Dressler Consulting, distributed training and multi-tenant AI platforms
- Ray API and head-cluster architectures for distributed training and inference, on both bare-metal Kubernetes with MetalLB and cloud-managed environments.
- Pulumi modules in Go for a multi-tenant AI platform: typed infrastructure code giving per-tenant Kubernetes, GPU pools and MLOps stacks, with automated teardown.
- An agentic harness for training-pipeline orchestration: subagent failure recovery, dynamic branching for hyperparameter sweeps, per-model-family skill loaders, and token-optimised calls for long-running jobs.
- Owned production alerting for distributed training across three and a half years of client engagements: GPU utilization, cluster health, cost anomalies and silent-failure detection.
Aceso Analytics, healthcare IoT
- Led the engineering team on a healthcare IoT platform serving LoRaWAN device fleets across multiple elderly-care facilities, with HIPAA-aligned operational practice.
- Architected the IoT and LoRaWAN stack on Kubernetes, with monitoring, alerting and edge inference for real-time device fleets.
- Mentored junior engineers and ran technical reviews.
Earlier: Afiniti, Halfpenny Software, Emumba
- Automated build, packaging and release for 40 or more microservices across air-gapped enterprise environments.
- Migrated legacy infrastructure to on-prem Kubernetes with production Ansible playbooks and Jenkins shared libraries.
- Ran parallel engagements for US and international clients on AWS, Azure and GCP: containerisation and clustering migrations with Terraform, Ansible and Bash, plus monitoring dashboards and observability stacks for hybrid and on-prem customers.
03 What is not on this page
Most of the work above is under client ownership and stays that way. Where a figure needs a baseline to be meaningful, ask and you will get the baseline rather than the headline.
The research that underpins the GPU work is public: the 2021 GPU pooling thesis, including what it measured and what it did not.