Building the tools
to think things
through.

Brain began as a self-hosted orchestrator: a way to give an AI agent a task in an existing project, see its output, and retain a record of the result. That structure created a useful foundation for other kinds of work.

Study tools needed materials, attempts, and feedback. Research needed sources, relationships, and notes. A newspaper needed collection, editorial preparation, and a separate public edition. These became parts of the same environment, with domain-specific workflows and shared application infrastructure.

The project now carries the public name well.education. Its central idea is practical: bring information, the tools for working with it, and a record of what happened into a place that can keep improving.

A useful system helps you see both the result and the work that produced it.

The learning workspace combines authored guides, question banks, practice sessions, and review. It includes LSAT preparation, Spanish, and university coursework, with different grading and feedback rules for different kinds of material.

Practice, feedback, and assessment

Learn mode offers immediate feedback during practice. Recorded attempts preserve the learner’s choices and make later review possible. LSAT attempts can capture eliminated choices, flagged questions, timing, and distractor-trap classifications. Spanish exercises distinguish accepted answers from accent errors and open responses.

Assessment state matters. In the LSAT workflow, an attempt taken with feedback revealed cannot establish a best score. In the communication-course workflow, question payloads omit answer keys and feedback commits the first selection. These choices keep instructional help and assessment records meaningfully separate.

Target the gap

The communication-course tool can build quizzes from weak concepts or selected trap tags, prioritizing missed questions while spreading concept coverage. Instructor-facing analytics group attempts by learner, concept, set, and distractor. These are explicit selection and aggregation rules, so the basis of a targeted quiz can be understood.

Give concepts a visual structure

Interactive lessons include worked examples, phrase-role annotation, comprehension checks, and immediate or deferred feedback. A reusable visual renderer supports argument maps, chains, comparisons, tables, and trees. Specifications are validated before rendering, including required fields and answer configurations.

News Unfiltered is the platform’s public editorial project. It brings source-linked reporting into a newspaper-style reading experience, with a workflow for gathering material, preparing summaries, and publishing an edition.

The editorial design distinguishes the reporting in a source from the interpretation built around it. AI assists preparation, while attribution, factual phrasing, and uncertainty remain part of the review. Each published summary links back to the reporting used.

Public publication is a separate surface from the private research workspace. That separation lets the project present a finished artifact without opening the underlying personal work environment.

The Neural workspace models research as entities, relationships, sources, excerpts, and claims. Evidence can support or oppose a claim, and review states give the resulting material a place in a continuing process.

Saved graphs offer an overview; entity profiles and evidence records let the reader move into specifics. Notes and human review changes keep the map connected to the judgment behind it.

The extraction path checks quoted text against its source passage. Quotations that cannot be found are discarded; matched wording is retained with its source context. This is a concrete validation step that makes generated research easier to inspect.

The original orchestration model has three parts. A project identifies the working context. An agent defines a reusable role and execution configuration. A task contains the requested work and produces a sequence of events and an outcome.

A separate worker can claim queued jobs, renew its lease, execute in the project’s environment, and stream events back to the server. This lets the web control plane stay on an always-on host while work executes where the necessary files and tools are available.

For local model runs, the background task is separate from the stream watcher. A viewer can reconnect and replay events without owning the task’s lifetime. Worker leases also provide a basis for bounded retries or recorded failure when execution stops responding.

Those details are part of making AI a usable application component: instructions enter a defined context, execution produces observable state, and a result becomes a record that can be reviewed.

The core is a Python/FastAPI application with a custom HTML, CSS, and JavaScript interface. Domain-specific modules support research, local AI, news, and other workflows. The system combines relational state with file-based course content and browser-local lesson progress.

This is an evolving application. Some workflows remain course-specific and some logic is concentrated in the central backend and frontend. The next architectural work is about making those proven patterns easier to reuse as the project grows.

Preserve work across change

Lesson progress uses fingerprints for individual lessons and migrates older browser stores. A content update can invalidate the affected lesson’s progress without discarding unrelated work. Course manifests can resolve question IDs against a shared bank so the same authored question can participate in different sets.

Separate execution from observation

Streaming output is valuable, but the open browser should not be the thing keeping a job alive. Separate task ownership, replayable events, worker leases, and recorded outcomes give execution a lifecycle beyond an individual connection.

Keep access boundaries explicit

Role-based API rules govern access to the private workspace. Public pages and published artifacts have their own surfaces. Public-web retrieval includes destination checks, and regression tests exercise rejected destinations and connection behavior.

Use the same visual language across tools

Thin Jost typography, a restrained white palette, translucent surfaces, and independently flowing, full-spectrum heatmaps give the different workflows a shared identity. The public presentation retains that language while giving the project more space to explain itself.

The wordmark moves from well.education to well.educated as the reader scrolls down, then returns on an upward scroll. It is a small expression of the larger idea: what you take in should change what you can do.

Thomas Holbird directs the project’s product design, implementation, and operation, using AI-assisted engineering throughout development. The work involves defining the behavior, connecting the parts, reviewing the result, and returning to the failures and rough edges that real use reveals.

The codebase demonstrates work across interface design, Python services, data modeling, asynchronous execution, local model integration, native app packaging, deployment, and regression testing. Its most useful story is the relationship between those disciplines: a feature has to fit the interface, the data, the execution environment, and the way someone will actually use it.

Current expansion work explores more reusable course infrastructure and experimental teaching workflows. These are directions for development, distinct from the implemented systems described above.