Ziloss Technologies · Salt Lake City

An AI‑operated company.
One human on payroll.

Ziloss automates small and growth‑stage businesses with custom agents, fine‑tuned open‑source models, and soon our own specialized models. The same stack takes on hard problems, like finding the next horizontal drilling site.

Agentsoperating
Human employees1
Inferencelocal + cloud
WorkstationM3 Ultra · 512 GB
Open sourceZBOT v1.0 · Apache‑2.0
Code written byagents, reviewed by Jeremy
24,961
Permian Basin wells scored for refrac and drilling potential
1
Human on payroll. Everything else is an agent.
512 GB
Unified memory for training and running models on‑site
$0.04
Nightly cost of ZBOT’s proactive memory batch
Services

Automation, priced like infrastructure.

Pick the layer you need. Every engagement ships as running software you own, with the model weights, the code, and the deployment in your accounts.

Agents

Custom Agents

Purpose‑built agents that run a workflow end to end: intake, follow‑up, scheduling, reporting.

ChannelsSMS · email · voice · web
Brainmulti‑model router
MemoryPostgres + pgvector
DeployCloud Run or on‑prem
Handoffhuman in the loop
Scoped quotefixed monthly after launch
Configure →
Models

Fine‑tuned Open Source

Open‑weight models trained on your data and domain, run where you choose. No per‑token lock‑in.

BasesQwen · Llama · Mixtral
MethodLoRA on MLX / CUDA
Evalvs. base, not just loss
ServingOllama · vLLM
Weightsyours
Scoped quoteper training run
Configure →
Coming soon

Ziloss Specialized Models

Our own models for narrow, high‑value jobs: lead qualification, operations, and geoscience screening.

Domainssales ops · energy
Footprintruns on one workstation
AccessAPI + weights license
Statustraining
Waitlistopen
Join waitlistearly access pricing
Notify me →
Operations

Business Automation

We take an existing process, instrument it, and replace the manual steps with agents and integrations.

Fits2–200 people
ConnectsCRM · sheets · calendars
ComplianceA2P · opt‑out · audit log
Timelinefirst run in 2–4 weeks
Supportagents, 24/7
Scoped quotefrom a one‑week audit
Configure →
Case study · Energy

Finding the next horizontal drilling site.

24,961 Permian Basin wells, ranked by machine.

For Permian Basin Explorations, Ziloss built a screening pipeline that ingests public FracFocus completion records, engineers features per well, and scores each one with a gradient‑boosted model for refrac and offset‑drilling potential. Leadership reviews candidates in an interactive map and table instead of a spreadsheet.

ClientPermian Basin Explorations
DataFracFocus · public filings
ModelGradientBoosting (scikit‑learn)
Wells scored24,961
Delivery4‑tab web app, dark theme
Outcomeranked refrac candidates for leadership
Score distribution · 24,961 wellsillustrative bins
0.00.51.0 wells top decile → field review
Models & infrastructure

Trained here. Served where you need it.

We fine‑tune open‑weight models on a single large‑memory workstation and deploy to your cloud or ours. Costs stay predictable because the heavy lifting runs on hardware we already own.

Training stack

LoRA fine‑tuning on Apple MLX against 32B‑class bases, with every adapter benchmarked against the base model for quality and speed before it ships.

Mac Studio M3 Ultra→512 GB unified→MLX LoRA→Qwen2.5‑Coder‑32B→eval vs. base
  • Datasets built per domain (the current Rust set is ~6,900 curated pairs)
  • Adapters versioned; contaminated runs are discarded, not patched
  • Local serving through Ollama for zero‑egress workloads

Production stack

Agents run as small services on Google Cloud Run with Postgres and row‑level security, a worker heartbeat, and secrets in Secret Manager. Nothing is hand‑deployed.

FastAPI→Cloud Run→Postgres 16 + RLS→Secret Manager→Prometheus · Grafana · Loki
  • Cost‑aware routing across Anthropic, OpenAI, Google, DeepSeek, and local models
  • Provider webhooks verified, idempotent, and replay‑safe
  • Right‑sized by default: one project’s monthly bill dropped 58% after a single review
Open source

ZBOT, the agent we run ourselves.

A Go‑based AI agent with a brain modeled on regions rather than one prompt. Released under Apache‑2.0 so you can read exactly how a Ziloss agent thinks before you trust one with your business.

Brain architecture

Router→Planner committee→Cortex→Thalamus→Hypothalamus sentinel

Fast models route and plan, a stronger model does the thinking, and a sentinel watches for drift. Each region can be swapped for a different provider or a local model.

Release

Repogithub.com/ziloss‑tech/zbot
LicenseApache‑2.0
Versionv1.0.0 · CI green
Imageghcr.io · Docker
Memoryproactive nightly batch, $0.04
View on GitHub
Company

How a one‑person company ships like a team.

Ziloss is operated by its own agents. They write the code, run the deployments, monitor production, and draft the proposals. One human sets direction, reviews the work, and signs the contracts.

That is not a gimmick. It is the product. Every client engagement uses the same agents and the same practices we run on ourselves, which is why a small business can afford them and a growth‑stage company can rely on them.

Jeremy LerwickCo‑founder · direction, review, client relationshipsHuman
Engineering agentsBuild, test, and deploy across GCP and local hardwareAgent
Operations agentsMonitoring, cost review, incident responseAgent
Research agentsModel training, evaluation, and data pipelinesAgent
Client agentsDeployed into customer businesses under their accountsAgent
Contact

Tell us the process you want gone.

A one‑week audit is usually the first step. You get a written scope, a fixed price, and a working first run in two to four weeks.

Emailjeremylerwick@gmail.com
Phone949‑230‑0036
Based inSalt Lake City, Utah