Audrey overview
This page is for the CEO, product managers, product and operations leaders, and technically curious stakeholders. No engineering background is needed.
It describes what the Audrey code does today. Engineers who need architecture detail should read the companion pages in the Audrey repository. Those pages are not copied here.
Audrey at a glance
Section titled “Audrey at a glance”What it is. The AI assistant inside the MarcoPolo app, for Pre-K and Kindergarten educators, and for administrators when that experience is enabled.
What makes it trustworthy. The curated MarcoPolo catalog, not the AI model, decides which resources exist and can be shown.
What it does not do. Write to teachers’ accounts, browse the web, or learn from conversations.
flowchart LR app["MarcoPolo app"] --> audrey["Audrey"] audrey --> content["MarcoPolo content"] content --> audrey audrey --> answer["Answer that cites that content"]
What Audrey is
Section titled “What Audrey is”Audrey is the AI assistant inside the MarcoPolo application. Its users are early childhood educators: mostly Pre-K and Kindergarten teachers working with children aged 3 to 5, and, when the feature is enabled, school and district administrators.
Teachers use Audrey to:
- get ideas, explanations, and coaching about teaching young children
- find MarcoPolo videos, educator guides, printables, and activities for a topic or for a unit of their curriculum (a study, such as a study about Water)
- plan lessons, including day-by-day plans built from a study’s materials
- ask how content relates to early learning standards (for example the Florida Early Learning and Developmental Standards, or the Head Start Early Learning Outcomes Framework)
- come back to a conversation and continue where they left off
Audrey is not a single AI model with a chat window on top. The language model writes the final wording of most answers, but a large part of what a teacher sees is decided by the rest of the system: who is asking, which curriculum they teach, what the curated MarcoPolo catalog says belongs to a study, what the teacher said earlier, which resources may be shown as cards, and which rules apply to the answer before it is published.
The most useful mental model: Audrey is an orchestrated system around language models. It combines curriculum context, a searchable content library, a curated catalog that is the authority on resources, conversation memory, safety and publication policies, and integration with the MarcoPolo application.
What happens when a user asks something
Section titled “What happens when a user asks something”Every question goes through the same broad sequence. Some questions finish early, such as a greeting or a request that needs clarifying first. Some take extra steps, such as a lesson plan or a list of all videos in a study. The backbone is this:
flowchart TD
A["Teacher types a question in the MarcoPolo app"] --> B["The app sends it to Audrey with the district, user, and conversation"]
B --> C["Load the conversation so far and the educator profile"]
C --> D["Decide what kind of request this is"]
D --> E{"Can it be answered without a search?"}
E -->|"greeting, clarification, or a question about the conversation"| R["Short answer"]
E -->|no| G["Work out the curriculum context: which study, which topic"]
G --> H["Search the content library or the curated catalog for this study"]
H --> I["Assemble instructions, profile, search results, and recent conversation"]
I --> J["Language model writes the answer"]
J --> K["Checks: which resources may be shown, and publication rules for administrators"]
K --> L["Answer appears on the screen; resource cards arrive at the end"]
L --> M["Conversation is saved for next time"]
R --> M
| Step | What happens |
|---|---|
| 1. Identify who is asking | The MarcoPolo application tells Audrey the district, the user, and the conversation. |
| 2. Understand the context | Audrey loads the earlier turns of the conversation and the educator profile the teacher filled in during onboarding. |
| 3. Decide what kind of request it is | A greeting, a factual question, an activity request, a request for videos or printables, lesson planning, a question about what was discussed before, and so on. |
| 4. Retrieve relevant material | Audrey searches MarcoPolo’s content library. When the teacher is working inside a study that has curated data, Audrey uses the curated catalog for that study instead. |
| 5. Use memory where it helps | Recent turns and a small index of topics the teacher has discussed are included so the answer follows the conversation. |
| 6. Generate the answer | A language model writes the reply from all of the above. |
| 7. Validate the result | Audrey decides which resources may appear as cards, removes things it is not allowed to show, and for administrators runs a publication check on regulatory claims. |
| 8. Return it | Text appears progressively in the app. The resource cards, and buttons where applicable, arrive when the answer is complete. |
| 9. Remember | The turn is saved so the next question can build on it. |
How Audrey understands context
Section titled “How Audrey understands context”Audrey builds a picture of the situation from several sources. None of them is guessed by the model. Each comes from a defined place.
| Context | Where it comes from | What it changes |
|---|---|---|
| District and user | Sent by the MarcoPolo application with every request | Whose memory and profile Audrey reads and writes |
| Session | Sent by MarcoPolo; one id per conversation | Which conversation the turn belongs to |
| Educator profile | Filled in by the teacher during onboarding in MarcoPolo and sent once to Audrey | Role (teacher or administrator), curriculum, age group, experience, students’ languages, state, and goals. Used to personalize the answer and to choose the curriculum context |
| Curriculum and study | The curriculum comes from the profile. The study is worked out from what the teacher says now and, for follow-ups, from the recent conversation | Which part of the catalog is relevant, and whether curated data applies |
| Previous conversation | Saved turns of this session | Lets “what materials do you have for this?” mean “for the study we were just discussing” |
| Earlier sessions | A small index of topics discussed in recent past sessions | Gives continuity when a teacher starts a new conversation |
A follow-up inherits context. A new self-contained request does not.
If a teacher talks about the Water study and then asks “what materials do you have for this?”, Audrey keeps the Water study. If the teacher then asks “show me videos about feelings”, Audrey does not quietly keep Water just because the new question names no study. Both halves of this rule are written down as product contracts and tested.
Audrey also receives a short block of onboarding context that the MarcoPolo user interface appends to some questions, for example the teacher’s curriculum and state. Audrey separates that block from the teacher’s own words before it decides anything, so the decision is made on what the teacher actually asked.
Teacher and administrator experiences
Section titled “Teacher and administrator experiences”Audrey has one core assistant and an administrator persona layered on top of it. Whether a user is an administrator is read from the role in their educator profile, for example “administrator”, “school_admin”, or “district_admin”. It is not taken from anything the user types. The administrator experience is only active when that feature is switched on.
What changes for an administrator:
- Audience. Audrey is told it is talking to an administrator. Topics such as budgets, staffing, HR, compliance, and operations are treated as part of the job, not as off-topic.
- Wording. Many fixed texts, including greetings, clarifying questions, memory summaries, notes about standards, and list endings, have administrator versions.
- Profile use. The classroom-oriented parts of the profile, such as curriculum alignment, children’s age group, and teaching experience, are not used to frame an administrator’s answer.
- Publication check on regulatory claims. An administrator’s answer may not state as fact what a law, regulation, licensing body, or external standard requires unless Audrey has an authorized source for it. A separate model reviews the finished answer. If it finds such a claim, or if the review fails, the whole answer is replaced by a fixed message that offers to help structure the decision and recommends confirming requirements with the responsible office. In that case no resource cards are shown.
- Delivery. Because the answer must be reviewed before anyone sees it, an administrator’s answer is delivered in one piece rather than word by word.
What stays the same: the content search, the curated catalog, the rules about resource cards, memory, and the request-type detection are shared between both roles.
While the administrator feature is on, the publication check is also applied to any user whose identity or profile could not be loaded. That user is not treated as an administrator. Audrey uses the base assistant, and still reviews the answer before publishing it.
Curriculum and trusted resources
Section titled “Curriculum and trusted resources”This is the most important product concept in Audrey.
The content library is a searchable library of MarcoPolo material: video transcripts, educator guides, activities, printables, alignment documents, and standards documents. Audrey searches it on most questions and uses what it finds as the factual basis of the answer.
The curated catalog is curated by MarcoPolo’s curriculum team. For each curriculum and study it records which videos belong to the study and how well each one matches the study’s concepts. Audrey loads a specific, versioned, and checksum-verified copy of this dataset.
When a teacher’s question is about a study that has curated data, Audrey switches to a curated turn:
- It selects up to six matching videos from that study using the curated ranking, and gathers their guides, activities, and printables.
- It builds a list of the resources allowed for this turn and tells the model which ones are available, by name.
- After the model answers, every resource card is checked against that list. A card for something that is not on the list for this turn is removed, even if the model mentioned it.
- When the model looks up “all videos for a study”, the answer is limited to the study Audrey resolved. A request about a different study is refused or corrected to the right one.
- Counts such as “this study has 14 videos” come from the curated catalog, not from the model.
The model may describe and recommend. It cannot make a resource exist, belong to a study, or appear as a card when the curated data does not allow it.
For curricula or studies without curated data, or when the curated data cannot be loaded, Audrey uses the general content library search. Cards are still built only from titles that match the real MarcoPolo catalog. A title that the model writes but that Audrey cannot find in the catalog does not become a card.
A language model can produce a plausible title for a video that does not exist, or attach a real video to the wrong unit. For a teacher, a broken or wrong resource is worse than no resource. The curated catalog makes these decisions deterministic and auditable. When the content library has nothing relevant, Audrey says it does not have specific content for that request instead of inventing some.
How Audrey uses AI models
Section titled “How Audrey uses AI models”Several AI models take part in a single conversation, each for a narrow job. They are Anthropic Claude models served through Amazon Bedrock. Which model runs each job can be changed per environment.
| Job |
|---|
| Writing the answer |
| Deciding whether a question is within Audrey’s purpose, when that check is on |
| Reading a lesson plan to decide which videos it used |
| Reviewing administrator answers for regulatory claims |
| Checking that a suggested printable really matches the request |
| Classifying an ambiguous request, writing a clarifying question, summarizing the conversation, and extracting what the teacher prefers |
The main model never works alone. What it is given is decided by the rest of the system:
- Instructions. A long, versioned set of instructions describes Audrey’s role, tone, boundaries, and how to present resources.
- Profile. The teacher’s curriculum, age group, and experience shape the answer.
- Retrieved content. The answer is grounded in what the search returned for this question.
- Authority lists. On curated turns the model is told exactly which resources exist for this request.
- Turn-specific instructions. For example “list all of them”, “this is a selection, not the full list”, “answer in Spanish”, or “do not invent an activity”.
After the model writes, the system decides what is published. “The model said it” is never, on its own, the reason a resource appears.
Many decisions do not use a model at all. Recognizing a greeting, most request types, the study a teacher means, follow-up inheritance, list requests, and action requests are handled by fixed rules, which makes them predictable and testable.
Memory and continuity
Section titled “Memory and continuity”Audrey remembers at three levels.
| Level | What it means |
|---|---|
| This conversation | Every question and answer is saved. When the next question arrives, Audrey reads the recent turns and gives the most recent ones to the model, so it can understand “that one”, “more like this”, or “yes, do that”. |
| What the teacher talks about | In the background, Audrey extracts a short index of topics discussed, plus preferences and follow-ups. The topic index gives continuity, including into a new conversation. |
| Who the teacher is | The educator profile is stored once and read on every turn. |
Audrey also writes a rolling summary of each conversation every few turns, and deliberately does not show that summary to the model. Summaries are model-written prose and often name resources without any authority behind them, which would let unverified titles re-enter later answers. Continuity comes from the recent turns and the topic index instead.
A conversation survives across servers because Audrey runs as several copies, and two consecutive questions from the same teacher can be handled by different copies. Nothing the conversation depends on lives only inside one copy. The turns, the profile, and the conversation state are stored in a shared database, and every copy reads them fresh at the start of each question. Each copy keeps small, short-lived caches for speed. Some background work, such as extracting preferences, can be lost if a copy stops abruptly. The conversation itself is not.
Resources, cards, and printables
Section titled “Resources, cards, and printables”Besides text, an answer can carry resource cards: clickable items that the MarcoPolo app renders below the message. Card types include videos from the MarcoPolo catalog, educator guides and activities linked to a video, and printables, including flash cards and photocards, linked to a video.
| Stage | What happens |
|---|---|
| 1. Candidates | Videos found by the search that the answer actually cites; titles the answer names and that match the real catalog; resources from the curated selection; and resources that Audrey itself pairs with the answer, for example the video an activity is based on. |
| 2. Validation | On curated turns, each candidate is checked against the curated list for this turn. Printable suggestions may be checked by a small verifier model when that path is enabled. Cards for things the answer itself says it cannot provide are dropped. |
| 3. Presentation | Cards are ordered by when the answer first mentions them, and duplicates are removed. |
Each card records why it is there: whether it came from a title the answer named, or from Audrey’s own pairing. Cards that Audrey added itself, rather than ones the answer named, are marked as related, so the app can present them as “related” rather than “recommended”.
Two special cases:
- “All of them” requests. When a teacher asks for every video or printable in a study, Audrey can answer with a complete text list, with counts taken from the curated catalog, instead of dozens of cards. On those turns no cards are shown.
- Low-confidence searches. When the search result is weak, Audrey does not show cards.
Actions
Section titled “Actions”Teachers sometimes ask Audrey to do something with a resource: save it to their plans, send it home to families, or use it in class.
Audrey recognizes these requests in English with fixed rules. If the request is ambiguous, with no clear resource or two different actions, it asks a short clarifying question. If the request is clear and names a resource from this turn’s cards, Audrey writes a short acknowledgment and the app shows action buttons under the answer. The MarcoPolo application performs the action when the teacher presses the button. Audrey’s text is told explicitly not to claim that it already saved or sent anything.
A direct integration in which Audrey itself calls the MarcoPolo application to save to “My Materials”, schedule for a classroom, or send home, with a two-step confirmation, exists in the code and is not active. It is accurate to say that Audrey can propose and prepare actions, and that the MarcoPolo application executes them. It is not accurate to say that Audrey writes to a teacher’s MarcoPolo account.
Safety, scope, and quality controls
Section titled “Safety, scope, and quality controls”Several independent layers protect the answer. Some can stop a request early. Some can replace or trim an answer that the model has already written.
| Layer | What it does | What the user sees when it acts |
|---|---|---|
| Request acceptance | Only requests that come from the MarcoPolo application are accepted | The request is rejected |
| Scope check | A model classifies whether the request is within Audrey’s purpose, including a safety category. The check can be switched off per environment | A fixed redirect message. If the check itself fails, Audrey refuses rather than answering unchecked |
| Retrieval authority | On curated turns, only resources from the curated selection can be offered or shown | Missing or corrected cards |
| Sufficiency | If the search finds nothing relevant, Audrey says it does not have specific content instead of inventing | An honest “no specific content” answer |
| Card validation | Cards must match the real catalog and, on curated turns, the curated list | Fewer cards |
| No invented links | Audrey does not add website links to answers | Links normally do not appear |
| Administrator publication check | Reviews administrator answers for unsupported regulatory claims | The whole answer is replaced by a fixed, safe message |
When something fails inside Audrey, for example the content library cannot be reached, the user gets a short message saying something on Audrey’s side is not working, rather than an answer built on missing data.
Streaming and the user experience
Section titled “Streaming and the user experience”Answers from the main model can take many seconds. The MarcoPolo app asks Audrey for a streaming answer: the text appears piece by piece as the model writes it.
The stream has a simple structure: a first event saying the search is complete, the text in many small pieces, and a final event with the resource cards, any action buttons, and technical details such as a trace id.
Consequences worth knowing:
- Cards appear at the end, after the text, because they depend on the final answer.
- Administrator answers arrive in one piece because they are reviewed before publication.
- Instant answers are slightly delayed on purpose. Greetings, clarifying questions, and similar short answers are held for about one second so the app’s live connection is ready to receive them.
What Audrey depends on
Section titled “What Audrey depends on”| Dependency | Role |
|---|---|
| Claude models on Amazon Bedrock | Writing answers and the narrow helper tasks above |
| The content library | Searching transcripts, guides, activities, printables, and standards |
| The MarcoPolo asset catalog | Looking up printables, flash cards, and other assets by video |
| The curated dataset | The authority on study membership and curated resources |
| The curriculum table | The list of studies and their videos and aliases for each curriculum |
| The memory database | Conversations, educator profiles, conversation state, and short-lived answer caches |
| The MarcoPolo application | Sends every request, owns the user interface, and executes action buttons |
If the AI models or the content library are unavailable, Audrey cannot answer normally. If the memory database is unavailable at startup, Audrey keeps running without memory. If a single read fails during a question, that question continues without memory.
What Audrey does not do
Section titled “What Audrey does not do”- It does not decide who the user is. The MarcoPolo application identifies the user. Audrey uses the identity the application sends.
- It does not invent resources on curated turns. Resources shown as cards must exist in the catalog and, on curated turns, in the curated selection.
- It does not write to MarcoPolo accounts today. Actions are proposed, and the application executes them.
- It does not browse the web or link to websites. Its sources are the MarcoPolo content library, the curated catalog, and the conversation.
- It does not state regulatory requirements as fact to administrators without an authorized source.
- It does not learn from conversations. Memory is stored per user and per conversation. Nothing in the code retrains or fine-tunes a model.
- It is not a general-purpose assistant where the scope check is enabled.
- The “parent mode” profile is marked as future in the configuration and is not a supported product surface.
Known behavioral caveats
Section titled “Known behavioral caveats”| Caveat | Detail |
|---|---|
| Administrator publication check | The administrator review can replace whole answers more often than intended in some request classes. This is an open functional area and is not addressed here. |
| Actions not certified end to end | See the Actions section. Audrey proposes actions. The application executes them. A direct write from Audrey to a teacher’s account is not active. |
| Languages | Audrey answers in the language the teacher writes in, and its detection recognizes eleven languages. Several fixed, system-owned texts are English only, including the scope-check messages, the administrator fallback message, and some list endings. Action requests are recognized in English only. “List all” grammar is recognized in English, Spanish, and Haitian Creole. |
One end-to-end example
Section titled “One end-to-end example”A Pre-K teacher whose profile names a curriculum with curated data is in the middle of a conversation about the Water study. The teacher asks: “Do you have any videos about floating and sinking for this?”
| Step | What Audrey does |
|---|---|
| 1. Identification | The MarcoPolo app sends the question with the district, user, and conversation. Audrey identifies who is asking and which conversation the turn belongs to. |
| 2. Context | Audrey loads the recent turns of this conversation and the teacher’s profile: role teacher, curriculum, age group 4 to 5. |
| 3. Request type | Fixed rules recognize a request for videos. It is not a greeting, not a memory question, and not an action, so no shortcut applies. |
| 4. Study | The question says “for this”, which is a referential follow-up. The question does not name a study itself, so Audrey looks at the recent turns, finds the Water study, and confirms it against the live list of studies for that curriculum. Had the teacher asked “show me videos about feelings”, Audrey would not have inherited Water. |
| 5. Curated selection | The Water study is in the active curated dataset, so this is a curated turn. Audrey ranks the study’s videos against “floating and sinking” using the curated ranking, selects up to six, and gathers their guides, activities, and assets. It now has a list of resources that are allowed on this turn, with the total number of videos in the study. |
| 6. Prompt | Audrey assembles the instructions, the teacher profile, the retrieved transcripts and guides, the list of available curated resources by name, the last few turns, and a note that the list is a selection rather than every video in the study. |
| 7. Generation | The main model writes a short answer recommending some of those videos, with a sentence on how to use each one in class. The text appears on the teacher’s screen as it is written. |
| 8. Validation and cards | When the model is done, Audrey extracts the titles it named, matches them to the catalog and to this turn’s curated list, removes anything not on the list, and builds the cards. If Audrey adds the video that a recommended guide or printable belongs to, that card is marked as related. |
| 9. Completion | The final event carries the cards, and action buttons where enabled. The app shows them below the answer. |
| 10. Memory | Audrey saves the question, a copy of the answer, and the titles it verified. Later, in the background, it updates the teacher’s topic index. If the teacher’s next question lands on a different Audrey server, that server reads the same saved conversation and the teacher notices no difference. |