pro
Jackie YU
AI agent engineer building systems that run unattended.
New to Contra
Jackie is ready for their next project!
My product. It continuously ingests the full Hacker News firehose and scores it for buying intent using some optimized open source model engines with a custom embedding pipeline — running on GPUs I operate, not API calls, so enrichment cost is amortized across the corpus instead of billed per request.
It runs unattended around the clock, and I publish the proof: the engine's live status page — corpus size, on-engine classify rate, real scan runs as they happen — is public at https://ferretout.ai/status.
0
58
Autonomous GPU fleet: 100 rented boxes, no daily human ops
An agent system that runs a GPU fleet on a spot-rental marketplace with no human in the loop day-to-day: it scans offers, places price-capped bids, installs and configures workloads on new machines, detects and heals dead or degraded instances, and tears capacity down under explicit destroy-gate rules — protected instances can never be touched, and hardware destroys always require human sign-off.
Spend is bounded by hard bid ceilings, so a bad market read can never blow the budget open. A live ops dashboard shows fleet state at any hour; I'm happy to walk through the real one on a call-free screen-share recording.
1
21
Open-weight LLMs in production: quantization, batching, failover
The infrastructure story behind FerretOut: what it actually takes to run open-weight models (27B-class) in production on rented GPUs. Quantization trade-offs measured against honest evals, batch scheduling that keeps the GPU saturated without starving interactive calls, and tunnel/failover plumbing so a single dead box never takes the product down.
The differentiator this unlocks: private inference on dedicated hardware — for HIPAA or data-residency-sensitive builds, your data never has to leave machines you control.
1
14
Voice companion agents for children & elderly — latency + guardrails
A voice-agent companionship system built for the two audiences where you cannot cut corners: children and elderly users. Sub-second conversational latency, guardrailed content behavior, and graceful degradation when the model or the network misbehaves.
If your product talks to vulnerable users, I've already made those mistakes on my own product so you don't have to make them on yours.
1
10