AI agents that only claim what they can defend.

Nikki is where the lab started: one agent on a Linux server with narrow, single-purpose skills for job hunting and real estate investing. Each skill was tested live on the server before the next one began.

• job-hunter: reads resumes, catalogues defendable skills, scans ATS boards, scores fit on an 80/20 bar, renders ATS-safe PDFs and tracks the pipeline.
• linkedin-magnet and serge-says: a recruiter-facing profile built from best-fit roles, and interview stories built only from real experience.
• deal-finder and area-finder: market scoring that runs free FRED checks first and spends on detailed reads only for the markets that pass.
• The Killington incident: Nikki once invented a confident scorecard instead of running the script. I redesigned around it so the agent judges and the scripts keep the record. That rule governs every later build.

Avow is the job-hunt stack sold as a product: an agent that finds roles you're genuinely strong for and never writes a claim you can't back up in an interview. I wrote the decision doc, chose the launch path, built the brand and put it live.

• Decision doc: weighed a hands-on MVP against a container-per-user platform and shipped the MVP first, so early revenue and feedback would steer the automation.
• Payments and access on Whop: no login system or card handling to build. Founding pricing of $29 a month or $79 once.
• Unit economics: modeled AI cost per member (about $5 a month typical, $16 heavy) against price, and found the one-time plan stops paying for itself after about 15 months.
• This site: a private S3 bucket behind CloudFront, ACM and Route 53, with strict security headers and no JavaScript, all in Terraform for about $1 a month.

Every Python script from job-hunter became a tool inside one n8n workflow. The model maps job requirements to skills, and a Code node does the math. The model can never hand a tool a score.

• 36offline tests run against the exact workflow file
• 1,770scoring cases, all identical to the Python original
• 16agent tools, shipped as one importable file

• One file to import: every tool calls back into the same workflow, so there's nothing to re-link after import.
• A real build process: each Code node lives in its own file, an assembler regenerates the workflow, and tests run on exactly what gets imported.
• Guardrails in code: a job enters the pipeline only with a score computed for that exact link, and a job board is saved only after a live fetch proves it exists.

Next step: the steps that need real credentials (model calls, Telegram, Google Drive) haven't run live yet. The README lists the seven checks for that first run.

I replaced the container-per-user plan with a serverless design where an idle member costs nothing. Members use the agent inside Whop, with Telegram as an optional second channel. The modules run in order: foundation, sensor, pipeline, reporting, compliance, app.

• 208resources in a clean Terraform plan
• 65tests running every Lambda against simulated AWS
• ~$62estimated monthly cost at 10 members

• Tenant isolation in IAM, not app code: no function can touch member tables directly. Each one assumes a role tagged with one member's id, and the policy only allows that member's keys. A bug gets AccessDenied instead of someone else's resume.
• Edge security: WAF in front of API Gateway, Whop's signed user token checked on every request, and signed webhooks with replay protection.
• Queues for scale: scanning, scoring and drafting run through SQS with concurrency limits, dead-letter queues and a stated ceiling tied to account quotas.
• A compliance gate: every claim in a generated resume is checked against the member's own documents before delivery. If one fails, the files are deleted and the reasons are logged.

Next step: planned and tested against a simulated AWS account. Real model calls, Whop traffic and the live isolation check come next.

Dottie reuses the Avow pattern for a harder customer: the broker pays, the agents use it, and neither may see what they shouldn't. It finds seller leads, judges markets, keeps past clients warm and builds listing-appointment packs.

• 103offline tests across every module and function
• 2 wallsagent from agent, and broker from agent, both in IAM
• 4 rule setsfair housing, sourced numbers, do-not-call, disclosure

• Two levels of tenancy: two agents in one office can't read each other's leads, and the broker sees team counts and shared leads only.
• A regulatory gate: blocks steering language, unsourced numbers, uncleared phone numbers and missing brokerage disclosure before an agent ever sees a draft.
• Reuse without rewriting: market scoring runs the original script byte for byte, with a test that proves it.

Next step: the Terraform was cross-checked but not yet planned against AWS. Running validate and plan comes first.