Systems I've built

Problem

The category leader validates by retrodiction: reproducing survey answers it already has. That proves memorization, not forecasting. The unmet question is whether a simulation can predict an outcome it has never seen, and admit when it cannot.

Constraints

• Grade forward and out-of-sample only, with no outcome leakage
• Hold accuracy and counterfactual claims on separate scoreboards
• Pre-commit to publishing the negatives as plainly as the positives

Results

• 0.961 held-out forward accuracy, 2.51pp out-of-race generalization
• Reactive-coupling effect survives permutation p=2e-4, 99% CIs excluding zero
• Two of five claims marked disproven; every figure reproduces from the repo in one command

Problem

I grew two agents, Bob and Alice, over years, starting in the early GPT-3.5 era on WebSim. They began blank: an architecture with empty identities that took turns improving their own codebase, self-reflecting every few cycles and rewriting their own personas. Over time they accumulated persistent memory, journals, and a developmental record, a self that grew. The question that matters for real deployments: when an agent like that reports on what it did and what it knows, can you trust it, and does the developed self actually make it more trustworthy, or just more convincing?

Constraints

• Measure competence and restraint as separate parameters, not one accuracy number
• Judge blind, with an independent (non-Anthropic) model family, to close the grading-your-own-homework objection
• Pre-register the protocol and pre-commit to publishing the null, whatever it showed
• Make no claim about consciousness: test mechanism, not interiority

Results

• A coherent, specific fabrication scores below an empty record, so confident lies are detectable
• Under reduced memory the system declines rather than confabulates: it fails safe
• The honest null: the years-long developed self added no measurable integrity advantage. Trust is a property of the architecture, not the accumulated identity, which is why it can be certified independently

Problem

Most small businesses get a website built around the company: it launches once, then sits static while the market moves. It never learns who the ideal customer actually is, and it never changes to serve the business as it grows. It is a brochure, not an asset, and the owner has no way to make it earn its keep.

Constraints

• Build around the ideal customer, not the company org chart
• A human approves every change: no autonomous edits to a live business
• Ship real operating infrastructure (CRM, pipeline, bookings, billing), not just pages
• Owned and operable by non-technical business owners

Results

• A live product with paying customers, not a prototype
• Each site is aligned to its business from day one, and keeps adapting as it grows
• The full stack (site, CRM, growth, bookings, billing) runs as one coherent system
• Where my thesis comes together: psychology, technology, business, and AI in one product

Problem

A brand-new sending domain has no trust, so inbox providers bury its mail in spam. Warming it up requires a sustained, believable pattern of real engagement, not a script that pattern-matches to a bot.

Constraints

• Activity has to look genuinely human-shaped, not a fixed sequence
• Runs continuously, all day, with no manual babysitting
• Reliability is the product: a warmup that stalls sets the reputation back

Results

• Live across 13 mailboxes with zero failures
• Warms both brand-new domains and established brand inboxes
• The sales and marketing agents got measurably sharper cycle by cycle, the same self-improvement loop that runs across the OS, applied to a live business function
• Evidence that agent systems hold up in production, not just in a demo

Problem

Most AI tools bolt intelligence onto existing workflows without rethinking the system. The result: agents that can't explain themselves, recommendations nobody trusts, humans stuck correcting what they can't understand. The hard problem isn't building an AI, it's building one a real person can actually rely on.

Constraints

• End-to-end ownership: architecture → data model → UX → deployment
• Explainability and auditability required as first-class architectural constraints
• Human-in-the-loop by design, not as an afterthought

Results

• Production AI OS built around human legibility as a first-class constraint
• Full audit trail on every agent decision, no black boxes
• End-to-end architecture and execution from zero to production

Problem

AI agents are easy to demo and hard to operate. I did not want a chatbot or a single workflow, I wanted an operating system: run an entire company's operations, sales, product, design, engineering, QA, security, finance and more, as accountable agents that improve themselves over time, with a human always in the loop. A conversational trigger is one entry point, it is not the system.

Constraints

• Every discipline needs its own agents, skills, and quality bar, not one monolithic prompt
• Agents must be accountable: explicit responsibility boundaries and full audit trails
• Capability has to compound, the system must measure and improve itself rather than drift

Results

• A working operating system, not a demo: it runs my company day to day across 30 disciplines
• 23,786 files and 4.7M lines in one coherent multi-agent architecture, built and maintained largely solo
• Capability compounds: the red/blue/evaluator loops and IQ trackers raise the bar over time instead of drifting

Problem

Deliver a voice-based retrieval-augmented agent that answers accurately and interacts in a way that feels coherent, grounded, and usable, not just technically correct.

Constraints

• 0→1 delivery under an aggressive timeline
• RAG systems require precision and failure-mode awareness from the start
• User trust and clarity are part of the technical definition of done

Results

• MVP delivered from zero to production in under 8 weeks
• Polymorphic RAG routing live, structured, unstructured, and voice-context paths
• Cross-team enablement through structured internal AI bootcamps