>the lab

This is the work I do outside the day job. The subjects are music, philosophy, and the Sierra, and the engineering discipline underneath them is the same one I use at work.

##featured-projects

##mastermind-alliance

An AI persona dialogue system with a hidden conversation director

An open-source platform for multi-persona philosophical dialogues. You pick two to five historical or philosophical figures, ask a question, and watch them debate it in their own voices, while an agent you never see steers the conversation toward the most interesting ground it can find.

// key features

Mastermind Tab: Multi-persona roundtable discussions with 2-5 historical/philosophical figures debating user questions
Conversation Director: A hidden agent that reads the transcript after every turn, selects which persona speaks next, and feeds them a leading question, optimizing for novel intellectual engagement. The objective is a prompt, so it can be pointed at any goal.
Backrooms Tab: Stream-of-consciousness AI dialogues exploring liminal spaces of thought
Persona Hierarchy: a dynamic persona inheritance tree built on CLAUDE.md-style context files, so a figure's voice, framework, and constraints compose from shared layers
Provider Abstraction: hot-swappable OpenAI, Anthropic, and OpenRouter backends behind one streaming interface

[01]

2–5 personas per roundtable

[02]

3 hot-swappable providers (OpenAI, Anthropic, OpenRouter)

[03]

Next.js 15 · React 19 · SSE streaming

// tech stack

Next.js 15React 19TypeScriptVercel AI SDKSSE StreamingAnthropic ClaudeOpenAI

impact

The part I keep coming back to is the director. Once the objective of a conversation is itself a prompt, the same machinery can be pointed at any goal, and a multi-agent dialogue stops being a fixed script and becomes something you can steer. It is the same idea I use in the compound pipeline's debate engine, where a plan has to survive an argument before it is allowed to execute.

$view repository explore

##wilsonic-autoresearch

Agent-in-the-outer-loop research over microtonal scale space

A Karpathy-style research harness pointed at an aesthetic domain. Erv Wilson spent his life searching for musical scales that serve melody and harmony at the same time, and this is an attempt to automate that search. A deterministic Python inner loop exhaustively scores scale families such as Combination Product Sets and Moment-of-Symmetry generators, and a Claude agent runs the outer loop, proposing hypotheses, writing pre-registrations before any code runs, executing batch sweeps, and filing verdicts in a public lab notebook. The subject matter is microtonality, and the engineering discipline behind it is the same one I use to make AI agents reliable in production.

// key features

Reward-Hacking Firewall: the scoring function is frozen, SHA-256-pinned, and CI-enforced, because an optimizer that is able to edit its own verifier is not really being verified at all
Goodhart's Law Caught Live: an early agent-proposed 1-cent generator maxed out the harmonic metric by cramming every tone within tolerance of everything else, and that failure is what motivated the freeze
Pre-Registration Discipline: hypotheses and predictions are written before the experiments run, and every run logs hypothesis → result → kept/reverted
Bit-Exact Cross-Validation: the offline Python scorer is validated against the compiled C++ plugin analyzer across 85 scales, with 0 mismatches
Exact Rational Arithmetic: the just-intonation paths use Python's fractions.Fraction so there is no floating-point drift in the ground truth, and the tempered paths carry a single hyperparameter (ε = 2¢)
Unit Tests vs. Folklore: melodic-axis tests for Rothenberg propriety, constant structure, and gap entropy caught errors that had been repeated for a long time, including the fact that the 12-EDO diatonic is not a constant structure
Falsifiable by Design: known limitations are documented in the open, critique is actively solicited, and every claim is graded by the strength of its receipt, because one good counterexample teaches me more than any amount of agreement

[01]

29/29 eikosanies land on the P = S diagonal — first a sweep result, then proven as a theorem

[02]

5 pre-registered experiments queued

[03]

Scorer frozen at v1.1.0, SHA-256-pinned in CI

// tech stack

Pythonfractions.FractionC++ / JUCEMTS-ESPClaude AgentMAP-ElitesCI Hash PinningScala .scl

impact

This showed me that the agentic-research pattern carries well beyond the SDLC. Frozen verifiers, pre-registration, and receipt-graded claims turn an aesthetic search problem into something you can falsify. I worked at Erv Wilson's side from 1995 to 2005, building his scale designs into software as he drew them, and this harness is a computational continuation of the search he ran by hand across thousands of diagrams. Decades of listening are what validate whatever the classifier finds.

$view repository $read the research explore

##explorations

Not everything I make is engineering infrastructure. This is what the same tooling produces when the deliverable is an educational interactive instead, with an original design system, dataset, interactions, and source discipline, built end to end with AI coding agents.

##ridge-receiver

The Sierra as a routing surface — an interactive field atlas for Tioga Pass

A ridge decides where water goes, ice can overrun that decision, and infrastructure can rewrite it altogether. This atlas traces how the Sierra routes ice, water, electricity, and people around Tioga Pass, from the Tioga-age glaciers that overran the crest, to the Lee Vining conduit that gives Mono Lake its one artificial outlet, to the morning climb out of a terminal-lake town straight up against gravity.

five layers, nine traces

Switchable ice / water / engineered / history / personal-route layers over a schematic oblique field model, with nine sender → payload → receiver traces and a date-controlled Mono Basin operating timeline.

every claim carries its receipt

An evidence drawer with 21 primary or authoritative sources. Where a claim is still only topographic inference I mark it visibly as unverified, and there is one published interpretation shown rejected and crossed out right on the map.

craft in the margins

Keyboard-accessible map nodes, reduced-motion mode, skip links, and a fully responsive layout, all inside a dependency-free static page that runs offline.

It is a different medium with the same epistemics. Like the research harness above, every claim here is graded by the strength of its receipt, so it is either cited, marked as inference, or rejected on the record.

$enter the surface skip to the map

##contact

// get in touch

[email]: marcushobbs@mac.com
[github]: github.com/marcus-w-hobbs
[linkedin]: linkedin.com/in/marcushobbs
[phone]: +1 (818) 802-4085
[loc]: Redondo Beach, CA

// open to roles and engagements building AI SDLC · enterprise engineering teams · AI labs