BeanGenie
An AI coffee discovery platform: crawls 1,700+ specialty roasters nightly and turns messy product pages into structured data with an LLM running on his own GPUs.
How it works →Infrastructure · AI · Chicago
Wynn Drahorad
Infrastructure engineer behind Tubi's production EKS clusters and its 100M+ logs/day OpenTelemetry pipeline. On-prem, he runs frontier open models at ~8,500 tok/s prefill, powering a crawler that indexes 28,000+ coffees nightly at a fraction of a cent per million tokens.
The system map
Click a stage to watch it run.
Ask about Wynn
Answered live by a model running on Wynn's own on-prem GPUs.
Answers come from a fact sheet Wynn approved, generated live by a model on his own GPUs. It can still be wrong.
Electricity only, measured at the wall. Doesn't include the cost of the GPUs.
By the numbers
100M+
logs/day through the OpenTelemetry pipeline Wynn led at Tubi
200+
nodes across Karpenter GPU/EFA node pools he provisioned
1,700+
coffee roasters crawled nightly by BeanGenie
28,000+
coffees indexed
~207tok/s
median decode and ~8,500 tok/s cold prefill, self-hosted GLM-5.3-Flash on 2× on-prem Blackwell GPUs
BeanGenie Inference Savings
$1,148 saved per month ($13,776 per year)
Product extraction, August 2026 · 100,974 LLM calls
Sonnet 5.5$1,155
100,974 × (3,712 in × $2/M + 401 out × $10/M)
On-prem$7
100,974 × (740 W × 2.19 s × $0.16/kWh)
2.19 s (one measured call) = 0.48 s prefill (3,712 in ÷ 7,662 tok/s) + 1.65 s decode (401 out ÷ 243 tok/s) + 0.06 s overhead
740 W = whole-machine wall power while busy, measured at the UPS
Extraction job only; other jobs are not counted and therefore real savings are even greater.
Now
Tubi
Senior Infrastructure Engineer, 2023–present- Automated creation of the main production and data EKS clusters end-to-end in a single quarter (Terraform, FluxCD, Vault, Karpenter).
- Evaluated Fluentd and Fluent Bit, recommended OpenTelemetry, and led the org-wide migration off the Datadog Agent across dev, staging and prod — 100M+ logs/day, zero-downtime cutover.
- Provisioned Karpenter GPU + EFA node pools for production data and AI/ML workloads running 200+ nodes.
- Built a vulnerability tracking dashboard that maps Wiz findings to owning teams.
Featured projects
Self-hosted inference lab
Two RTX PRO 6000 Blackwell GPUs serving frontier open models behind llama-swap (vLLM/SGLang), with every request traced in Langfuse. Wynn tunes precision, KV cache, speculative decoding and scheduling per model, and measures cost per token from wall power.
starfleet on-prem cluster
A multi-node k3s cluster run entirely through Flux GitOps: observability on SigNoz/ClickHouse, self-hosted services, CI runners, and the BeanGenie crawlers.
Experience
- TubiSenior Infrastructure Engineer
- BeanGenie LLCFounder
- Fetch RewardsSenior DevOps Architect
- Slalom ConsultingSenior DevOps Engineer
Toolbox
- Infrastructure
- Kubernetes, Terraform, AWS, FluxCD/ArgoCD, Karpenter, OpenTelemetry, Vault
- AI
- vLLM, SGLang, multi-model serving, tensor parallelism, speculative decoding, quantization
- Code
- Go, Python, TypeScript, SQL