Gabe

Marketing Engineer · Forward-Deployed AI Engineer

Hi, I'm Gabe. I build AI harnesses, tools, agents & systems.

Fifteen years in marketing taught me to ask why someone would care about a product. I've worked on positioning and spent years researching audiences. Ghostwriting meant learning how someone else thinks before writing a word for them.

For the past five years, I've applied that judgment to GTM engineering and AI systems. That taste shapes what I build and how I judge the output. I build the workflow with the people who will use it.

My working set includes Claude, x402, Python, SQL, RAGas, embeddings, vector databases, MCPs, agentic systems and harnesses, evals, context engineering, TypeScript, JavaScript, and the APIs around them.

How I think about work and AI: this video, this tweet, this doc, and this one on GTM.

Project	Stack	What it does
Tixmancer		AI agent that finds overlooked secondhand items and checks the asking price against your limit, working toward buying below market value.
Hedwig		Shared, revocable roles for related Solana apps: each app authenticates its users or agents and checks their access onchain.
scout-onchain		For onchain traders: explainable PnL, technical reads, and DCA entry alerts, built on the Zerion API.

Project	Stack	What it does
prospector		Ran 100 outbound leads in 48 hours for $4: two scrapers, Clay enrichment, ICP gates, human approval, fully forkable.
launch-factory		One release folder in, six launch assets and a campaign calendar out, every claim sourced, nothing ships without a human.
gtm-tam-mapper		Turns 75 raw rows into 40 deduped, tiered accounts, every disqualification named.
gtm-outbound-personalizer		Master doc plus account list in, per-tier variants out; QA blocks rows missing pain points.
gtm-objections-to-content		Turns 5 call transcripts into 12 tagged objections and a 2-week calendar; refuses happy calls.
gtm-campaign-launcher		Checks list, copy, and sequence in one preflight; broken campaigns emit zero configs.
gtm-content-repurposer		Indexes 18 posts, ranks the closest match by BM25, refuses off-topic queries.
gtm-campaign-analyzer		Flags the winning segment combo, 2.14x over tier average; refuses winners on small-sample weeks.

A harness I have designed and fine-tuned for months: HyperBots. 32 specialist agents run through one routing layer. Deterministic gates block a bad release before it deploys.

This is an operating layer with context budgeted per seat, routing covered by automated tests, verification gates, and a skills-and-hooks system that enforces its own rules.

→ Router. State the outcome; routing picks the specialist, tie-breaks pre-judged in writing.
→ Mechanical gates. A pre-ship gate blocks every push until a reviewer signs the exact commit. One push human owner.
→ Verification as code. Hash manifests diff source against the installed runtime and trigger eval suites to gate routing changes. CI catches drift before merge.

The same crew builds the feature and gates the ship. It also writes the launch copy and tracks whether AI engines cite it.

An augment is a folder that teaches an AI assistant a whole job. It is onboarding for an AI employee, plus guardrails that stop it from skipping steps, zipped so anyone can install it.

Every augment here is three parts:

→ Skills = the SOPs. Step-by-step playbooks the AI follows for each task.
→ Hooks = the guardrails. Automatic checks that block the AI the moment it breaks a rule.
→ README = the manual. A stranger installs and runs the whole thing without me in the room.

Four augments, each one shipped real work before it was listed:

Also shipped: kageb · patronus · worldcup-settlement · meteora-jup-sol-safety-monitor · solguard · mermail-defi-navigator · safeskill · wingman · clawmanship · agenthub · hermes-aria-theme · walletbrief-monad · yieldpilot · beat-claude-engineer-004-analytics-pipeline · ecommerce-sales-analysis · solana-narrative-tracker.

Security researcher, 4 bounties paid: a high-severity Anchor bug in LazyAccount (writeup, PoC, fix, upstream shipped the same fix five weeks later), a fund-lockup bug in Ern Protocol, and two Quantus findings, guardian enrollment without consent in set_high_security and scheduled transfers below the existential deposit stranding funds with no owner-accessible recovery. Other reports stay unpublished while under review, per program terms +++ also built the Solana Vault Standard Extension for a Superteam bounty.

I write about GTM engineering, all things marketing, web3, proof, evals, and opinions at my blog.

gabeonchain.com · X @gabe_onchain · LinkedIn