Michael Hoefert

Instead of relying on a single AI prompt, I have challenger and defender agents literally debating each other to fact-check claims and kill hallucinations before you ever see the output. I also integrated LangSmith so we can actually understand and measure agent performance. To keep the AI from hallucinating logic, I built a Shared Tool Registry—the agents are bound to deterministic Python scripts they can call to scrape job descriptions with Firecrawl, execute exact KNN vector searches against a Postgres database, and securely decrypt PII during document export.

But the coolest part is the Compounding Memory System. In the background, Joey's memory agents analyze your edits and feedback to build specific writing 'skills' tailored to different job titles, company sizes, and industries (e.g., a Staff PM at a Growth-Stage FinTech). When you apply for a job that fits those criteria, the system's writing agents automatically retrieve and use those exact skills via a HyDE (Hypothetical Document Embeddings) RAG model. These skills are constantly refined every time you provide feedback or make an edit, meaning Joey learns your unique voice over time. The product uses some of the best agentic memory practices so that you aren't constantly context seeding. Coupled with a 'Labor Illusion' UI that shows you the agents' thought processes in real-time, it's an engine built from the ground up to actually learn your career narrative and never lie.

Built over 7+ months as a solo founder project, the real complexity lives in the closed feedback loop the platform creates. Every graded call feeds a "maverick" detection layer that surfaces wins where reps broke the playbook and still closed deals. Each maverick is analyzed by complex context-seeking LLM APIs to distinguish replicable techniques from deal-specific context, then verified by the manager. The manager verification stage is crucial; it provides a human-in-the-loop verification gate but also allows managers to add commentary on why they are or aren't verifying a call. The manager commentary is then stored and used to power the self-learning maverick detection engine so that over time, the detections are more aligned with what the manager wants. Once enough verified mavericks accumulate, a correlation engine cross-references bypass frequency against historical coaching session records, and when a criterion has been coached repeatedly but top performers keep skipping it in winning calls, the system classifies it as terminally flawed and flags it for removal in the next evolution proposal. That proposal renders as a side-by-side rubric comparison with per-change rationale and evidence strength, with the pre-computed frequency analysis injected as structured context. The result is a methodology that updates itself from the team's own winning patterns, and a system that can tell a manager not just that their reps aren't following the script, but that the script is wrong.

The point is that I never have to seed context again. Whichever harness has the leading model that week, I point it at the vault and it already knows my work, my writing and my products.

It started as two separate systems. The AI Intelligence wiki ingests raw market signals (web clips, podcasts, newsletters), classifies them, and keeps a queryable map of frontier models, technologies and startups. The Career Expansion vault ran a 6-agent resume pipeline and a 5-agent Q&A compiler with an advocate-critic loop and a cold factual audit, and it became the blueprint for Joey. Now they share one vault, and the links between them are where the value is: 66 links tie my AI research to Joey's build docs. That's how I learned how Sierra built its compounding memory system and brought the same idea into Joey.

Built as a fully custom, self-improving training engine, the real complexity lives in the automated closed feedback loop and how the data flows into it. Every activity logged on Strava triggers a scheduled GitHub Actions workflow that streams raw sensor data into our Python math engine. Here, we run decay calculations to track CTL (Fitness), ATL (Fatigue), and TSB (Form), along with cardiac drift (Pa:HR) metrics to flag aerobic decoupling, and write the results back to Airtable. Airtable is also where my subjective recovery notes and IT-band pain logs live, right next to the ITB rehab research Gemini Deep Research pulled together for me.

Every morning, the coaching agent runs an analysis on the updated metrics, cross-references my recovery notes, pain logs, and the rehab research, and writes back detailed coaching narratives. Its coaching and workout-modification instructions live as plain markdown in my own repo rather than inside any one tool's automation feature, so I can point Claude Code, Codex, or Cursor at them without rebuilding anything. Raw data is server-side rendered directly from Airtable, ensuring the dashboard is always hydrated with live metrics, and the narratives actively shape future workout intensities, closing the loop from raw biometrics to physical adaptations.

The rule is deterministic first. Python does the routine work: scanning inboxes, fixing frontmatter and broken links, rebuilding indexes, and checking that key docs haven't gone stale. A single call to a small, cheap model only happens where judgement is needed, like classifying the day's AI news or turning build transcripts into spec updates. Jobs exit early when there's nothing new, so most runs cost nothing and the whole system runs on about $0.50 a month in LLM costs.

Every change lands as a pull request. A rules-based merger with a branch allowlist, path fences and its own tests decides what merges, so PRs from different agents and harnesses all go through the same gate. The newest job, Model watch, scans seven AI labs every Monday for new models, price changes and retirements, and proposes updates to Joey's model registry in one standing PR that I review and merge by hand.