Portfolio

Michael Goss

AI Automation Specialist · AI Operations · AI Agent & Workflow Automation

I weld to code by day. I build AI systems the same way: to spec, with safety controls, and honest about results.

Currently: running 24/7 radio ops for Spaces Media and refining the paper-trading research system

I design, deploy and run production AI agent systems for real small businesses, starting with my own. As a certified structural welder, I work to spec, put safety controls first, log everything, and report results honestly.

Richardton, ND michael@michaelgossimaging.com X: @Sir_Coopcelot
Download resume (PDF)
24/7live radio automation
25+shows a week
~133bots tracked publicly
37pre-registered strategies tested

Case studies

Each one is a live system: the problem, what I built, the controls, and what happened.

24/7 broadcast25+ shows a week, thousands of hours on air
~133 bots rankedpublic, after-fee, with a published methodology
37 strategies, no edgepre-registered and reported as-is
Approval-gated outreachverified contacts and a do-not-contact list
Spaces Media at a glance: 24/7 on air, 25+ weekly shows, 1000s of hours on air and about 5.4k followers on X, beside the station\'s microphone artwork
Result: 24/7 on air with 25+ weekly shows, per spaces.media; ~5.4k followers per X.

Spaces Media: 24/7 X Spaces radio automation Founder

spaces.media · @Spaces_MediaLLC

24/7on air
25+shows/week
1000shours on air
~5.4kX followers

An AI operations agent keeps a 24/7 Web3 radio station live on X Spaces and hands each show to the next host on time.

How it works
Problem
A Web3 radio station on X Spaces that runs around the clock with 25+ shows a week. Live Spaces drop, clients crash, and every show has to be handed to the next host on time. Doing that by hand day and night isn't sustainable.
What I built
  • An AI operations agent ("Orion", built on Hermes Agent) that manages the live Space on a Windows desktop, with Discord as its control and alert channel.
  • Keep-alive: it detects a lost broadcast, goes live again after 1 minute down, and retries every 2 minutes.
  • Crash recovery, plus show-to-show host handoffs.
  • Schedule management: a daily 6 a.m. schedule post, public warnings and host DMs 30 and 5 minutes before each show, and a title change for each show.
  • The Spaces Media website, built with AI coding agents.
Controls
  • A handoff counts as done only when the next host is confirmed as co-host in the room. A click that looks right isn't enough.
  • Routine health checks run as plain scripts with no AI model cost.
  • Dangerous commands need my approval, and credentials stay in a vault the agent can't read.
  • Every incident becomes a written procedure.
Handoff rule
Handoff rule diagram: step 1, next host invited; step 2, the guest list must show COHOST; step 3, handoff counted and logged. If not confirmed, retry, then escalate.
Results
The station broadcasts 24/7, runs 25+ shows a week, has thousands of hours on air, and has about 5.4k followers on X.
BotScoreboard.com: the full \'Read this first\' sampling-bias note above the top leaderboard rows, with Suspected and active labels and last-30-day, all-time and after-fee results
Result: ~133 bots ranked live, with the sampling-bias note shown before the leaderboard.

BotScoreboard.com: a public data product

botscoreboard.com · methodology

~133bots tracked
Publicmethodology

A live leaderboard of what Polymarket trading bots actually keep after fees, losers included.

How it works
Problem
Trading-bot sellers post screenshots, not audited results. Nobody showed what prediction-market bots actually keep after fees.
What I built
  • A live leaderboard of about 133 Polymarket trading bots, scored from each wallet's public on-chain history.
  • Results are after trading fees and include rebates. Losing and idle bots are listed too.
  • An automated data pipeline, static site hosting on Cloudflare Pages, a bot-submission form, and a newsletter signup.
Controls
A published methodology. Confidence labels (Confirmed / Self-claimed / Suspected). A note on sampling bias, because bots taken from "top bot" lists over-represent winners. Only public data is used.
Results
Live and public. Data last refreshed Oct 19, 2026.
Paper-trading pipeline diagram: signal, rules, risk checks, kill switch, paper order, decision log, then a daily review that proposes fixes
Result: 37 pre-registered strategies, none showed an edge, reported as-is.

A disciplined AI paper-trading research system

Alpaca paper-trading API · research and paper trading only, no real money

37pre-registered strategies
0showed an edge
Paperonly, no real money

An options research system with hard risk limits, built to test honestly for an edge before any money is at risk.

How it works
Problem
Test whether options strategies have a real edge, without fooling myself, before any money is at risk.
What I built
Working with AI coding agents, I built an end-to-end system: a weekly universe build, an options-flow scanner, defined-risk spread construction, scheduled trade management, end-of-day reports, and an automated daily review that proposes fixes. A daily pre-market research briefing written by a separate research agent feeds into it.
Controls
  • Paper account only, enforced in code.
  • A kill switch: a 6% drawdown halts new trades until I approve a restart, and there's a 2% daily-loss stop and a pause when data fails.
  • At most 20 orders a day, with no market orders and no uncovered shorts.
  • Every decision is logged with the reasons behind it.
  • Change control: risk limits can never be loosened, and parameters change at most once a day.
Results
37 strategies tested, all pre-registered, with sealed holdout data, out-of-sample splits, deflated Sharpe ratios and double-cost stress tests. None showed an edge, so the holdouts stayed sealed and I reported the result as-is instead of curve-fitting. Paper trading began Oct 19, 2026 to collect forward data. No profits are claimed.
Exterior 3D render at dusk: a lit rectangular pool with a fire bowl on the deck, palms and a lounge terrace along the back wall
Studio render: a pool and deck lit at dusk. Result: outreach for the studio runs only on verified contacts, with my sign-off.

Multi-agent operations for Michael Goss Imaging

michaelgossimaging.com · exterior 3D landscape and outdoor-living renderings, built in Structure Studios VIP 3D

Specialist agents handle research, outreach and records for a solo studio, with nothing sent without my approval.

How it works
Problem
I'm a solo operator who needs research, outreach, follow-up and record-keeping done without hiring staff, and without risking spam or sending the wrong thing.
What I built
  • An orchestrator agent that routes work to specialist agents for market research, lead research, web design, email and records.
  • A cloud agent linked to a self-hosted desktop email agent through two-way messaging and authenticated webhooks.
  • A daily operations brief.
Controls
  • Each agent has one job and written authority limits. Nothing is sent and no money is spent without my approval.
  • Contacts are verified on company websites only, never guessed.
  • Opt-out scanning, a bounce circuit breaker, a single do-not-contact list, and a no-marketing rule for existing clients.
Results
A compliant, logged outreach system with these controls built in:
  • Verified contacts only: each address is confirmed on the company's own website, never guessed or bulk-scraped.
  • Do-not-contact enforcement: a single list that every agent checks before drafting, queuing or sending.
  • Approval before sending: outbound mail goes out only after my sign-off.
  • Periodic audits: I review the outreach system's sends, bounces, opt-outs and configuration, and fixes ship only after I approve them.
More renders
Daytime render of the same pool with a fire bowl, sun loungers and palmsRender of a pool terrace with umbrellas, loungers, a raised stone planter and flowering plantsNight render of a stone patio with string lights, uplit trees and a pool edge

Skills

Evaluation & risk controls

Pre-registered testing, sealed holdouts, out-of-sample validation, deflated Sharpe ratio, cost stress tests, kill switches, order limits, decision logging, change control, deliverability compliance

AI agent operations

Multi-agent orchestration, agent role and prompt design, scheduled jobs and webhooks, agent-to-agent messaging, desktop GUI automation, human-in-the-loop approvals, runbooks, incident review

Integrations & tools

Grok Bot, Hermes Agent (Nous Research), Alpaca API, Slack, Gmail, Google Workspace, OneDrive, GitHub, Cloudflare Pages, Discord, X Spaces, MCP servers, Python (AI-assisted)

Business & trade

Structure Studios VIP 3D, exterior 3D rendering, client management, B2B outreach, AWS D1.1 certified welder (FCAW-G 3G/4G), MIG (GMAW)

Download resume (PDF)

How I work

The same habits I use in welding: work to code, check the work, and don't sign off on something you haven't inspected.

Spec first

Every system starts as a written specification: goals, inputs, success checks, and what the agent must never do.

Safety policies

Hard limits are written down and enforced in code: kill switches, caps, do-not-contact lists, and paper-only checks.

Human approval

Anything that sends, spends, or can't be undone waits for my OK.

Logging

Every decision and incident is recorded, so I can see why something happened and fix it.

Honest evaluation

Tests are pre-registered, holdouts stay sealed, and I report results as-is, including when the answer is "no edge."

AI-assisted development

I design the system and direct AI coding agents, then read, test and run the resulting code.

What I won't do

The lines my systems are built not to cross.

No profit claims from paper

Paper results stay paper. No real-money or profit claims.

No guessed contacts

Every contact is verified on the company's own site, never guessed or scraped.

No loosened limits

Risk limits are never loosened to force a trade.

No unapproved risky actions

Anything that sends, spends or can't be undone waits for my OK.

No secrets in agent hands

Credentials stay in a vault the agents can't read.

No unlogged changes

Nothing ships without a log of what changed and why.

The paper-trading daily loop

  1. Session startpaper-only and data checks
  2. Morning briefingpre-market research brief
  3. Entriessignals pass rules and risk checks
  4. Managementscheduled trade management
  5. Closeend-of-day report
  6. Daily reviewautomated review proposes fixes
  7. Next-day improvementsparameter changes, at most once a day