IDÉLLIA case study
RAGEdTechGenAI0-1Franco-Ontarian

IDÉLLIA: Building the AI Pedagogical Co-Pilot for Franco-Ontarian Education

How I designed a RAG-based instructional assistant that transforms 15,000 certified educational assets into curriculum-aligned “Learning Journeys” — in under 2 minutes.

Role

Product Manager

Organization

TFO (Télévision française de l'Ontario)

Platform

Idéllo · idello.org

Status

PRD Complete · In Development

Context & Problem

The Paradox of Abundance

The problem wasn’t a lack of content. With nearly 15,000 certified educational resources — documentary series, digital books, podcasts, pedagogical tools, and interactive games — Idéllo already hosted one of the most comprehensive Francophone educational libraries in Canada. The platform was purpose-built for Franco-Ontarian schools: ad-free, privacy-compliant, curriculum-aligned, and culturally specific to the Franco-Ontarian identity.

And yet, teachers weren’t using most of it. The platform’s “Advanced Search” put the entire cognitive burden on the user. To build a single lesson, a teacher had to search for a video, watch it to verify quality, search again for a supporting activity, and manually stitch those pieces into a coherent sequence — all while checking that each asset matched the specific Ontario curriculum strand for their grade level.

The irony was sharp: the more content Idéllo added, the harder discovery became.

Meet Gabrielle — The User Archetype

Grade 5 teacher · Francophone school in Ontario · Tech-literate but time-poor

She operates under strict school board mandates: no ads, no student logins, no data collection. She trusts Idéllo because it’s the “only public Francophone media entirely dedicated to the Franco-Ontarian community.” What she doesn’t trust is spending 45 minutes on a Sunday evening building a lesson that should take five.

Her core jobs-to-be-done

  • Curriculum compliance: She needs content aligned to the Ontario Life Systems strand for Grade 5 — in French, without Standard European French dominance.
  • Cultural validation: She seeks content reflecting a "plural Francophonie" — Acadian accents, Indigenous perspectives, Franco-Ontarian voices.
  • Instant differentiation: She needs to quickly adapt a lesson for a student reading below grade level without building a second lesson from scratch.
  • Frictionless deployment: She wants the right content on students' tablets right now, without requiring accounts or complex navigation.

The Five Pain Points

1

The Haystack Problem

Searching 15,000 resources through rigid metadata filters (Grade, Subject, Theme) creates decision fatigue. Teachers frequently abandoned the platform for Google or YouTube — faster tools that violated school board policies on safety and advertising.

2

The Stitching Tax

The platform treated assets as isolated files. A teacher who wanted a Hook → Instruction → Practice structure had to manually find and sequence all three pieces — a manual orchestration tax that prevented accessing the full depth of TFO’s multimedia ecosystem.

3

The Safety-Efficiency Paradox

Teachers were caught between two unacceptable options: fast but unsafe (ChatGPT, YouTube), or safe but slow (Idéllo’s manual search). No tool offered the speed of generative AI within a guaranteed, certified, ad-free environment.

4

The Cultural Blindspot

Standard keyword algorithms rank by popularity, systematically burying minority accents and Indigenous perspectives in favour of Standard European French. Surfacing culturally specific content required query sophistication most teachers didn’t have.

5

The Deployment Barrier

TFO’s strict privacy policy means student email addresses are never collected — making distributing curated content to students genuinely difficult. Without an anonymous access mechanism, teachers defaulted to projecting content passively rather than enabling individual student exploration.

Gabrielle, a culturally-conscious Francophone educator, spends too much time manually filtering 15,000 disjointed resources to build a curriculum-aligned lesson sequence, and she frequently fails to deploy safe, culturally validating learning journeys to her students because the only efficient alternatives are unsafe, ad-supported, or Anglophone platforms.

Opportunity & Market Insight

A Blue Ocean in Francophone EdTech

The Ontario educational technology landscape is bifurcated by language — and that divide creates a distinct market opportunity.

Anglophone Market

Robustly served

TVO offers a full instructional stack: TVO Learn (curated discovery), Mathify (live 1:1 tutoring), and mPower (gamified STEM) — serving millions of students.

Francophone Market

Structural absence

TFO’s Idéllo is the content library — but there is no active instructional layer, no AI-driven sequencing tool, no anonymous deployment mechanism at scale. This isn’t a feature gap. It’s a structural absence.

Competitor Analysis

PlatformStrengthsGaps
TVO MathifyAnonymous, human-led 1:1 tutoring; strong privacy modelHuman-in-the-loop = not scalable; English only
TVO LearnCurated discovery, curriculum-taggedStatic metadata search; no sequencing
ChatGPT / Open LLMsFast, fluent, powerfulHallucination risk; ad exposure; no curriculum grounding; privacy violations
Idéllo (current)Certified content, ad-free, culturally specificPassive library; cognitive burden on teacher; no AI layer

Key insight: the institutional competitors are deliberately avoiding generative AI. IDÉLLIA’s opportunity was to be the first platform to solve the safety-efficiency paradox at the institutional level.

Three Market Voids

Void 1 — Static vs. Dynamic

Teachers have a massive passive library but no active instructional partner. IDÉLLIA converts static assets into dynamic, sequenced Learning Journeys on demand.

Void 2 — Safety vs. Speed

No existing tool offers both. IDÉLLIA occupies the "Safe AI" gap by restricting all LLM retrieval to TFO's certified ecosystem — zero hallucination of external, unsafe content.

Void 3 — The Cultural Blindspot

Generic AI defaults to global English or Standard European French. No competitor — institutional or commercial — addresses the Franco-Ontarian community's specific cultural mandate.

Vision & Product Strategy

Transform Idéllo from a passive archive into an active pedagogical partner — where any teacher can generate a secure, curriculum-aligned, culturally inclusive Learning Journey in under 2 minutes, without exposing a single student data point.

This vision encodes three non-negotiable constraints that defined every product decision: Speed (under 2 min, not 5), Safety (zero student data exposure), and Cultural specificity (Franco-Ontarian French, not just French).

The Core Product: Parcours d’apprentissage (Learning Journeys)

The teacher’s input is a natural language prompt: “Create a lesson on Water Systems for Grade 5 focused on Indigenous knowledge.” The output is a fully sequenced Parcours, deployable via an anonymous Class Code.

Hook

An engaging video to open the lesson and capture attention

Instruction

A digital book or documentary to build knowledge

Application

A game, interactive tool, or pedagogical strategy card to reinforce learning

Extension

Optional differentiated activity for additional challenge or support

The Walled Garden Architecture

IDÉLLIA’s AI would be powerful but bounded — it could only retrieve content from TFO’s certified ecosystem. This wasn’t a limitation; it was the core value proposition. It simultaneously eliminated hallucination risk, satisfied school board privacy mandates by design, and made IDÉLLIA irreplaceable — no generic AI tool could replicate curriculum-aligned Franco-Ontarian content retrieval.

Feature Prioritization (RICE Framework)

FeatureReachImpactConfidenceEffortRICEPriority
Parcours Generator (NL prompt → sequence)90%1097115P0
Anonymous Class Code Deployment90%995145P0
Cultural & Identity Re-ranking70%87665P1
Differentiation Mode60%87748P1
Home-School Bridge ("Generate Home Link")50%76452P2

Data & AI

Why AI — and Why It’s Essential, Not Optional

Scale & Semantic Complexity

With 15,000 assets, rigid metadata filters create the Haystack Problem. A query like "Grade 5 lesson on conservation of energy with an Indigenous perspective" requires semantic understanding across multiple curriculum tags, cultural markers, and content formats simultaneously. Keyword search cannot do this.

Multimodal Orchestration

The ecosystem contains documentary series, audio podcasts, digital books, games, and pedagogical tool cards — all in separate silos. AI acts as connective tissue, linking these formats into a coherent instructional sequence.

Privacy & Cultural Mandate Compliance

Only AI can operationalize the Walled Garden constraint at scale — strictly restricting retrieval to TFO's certified catalog while simultaneously weighting culturally specific assets against global popularity signals.

The RAG Architecture

Why RAG over fine-tuning — a deliberate, strategic choice:

  • Source of truth stays in the catalog: As Idéllo's content library updates — new videos, new curriculum standards — the retrieval layer updates automatically. A fine-tuned model would require periodic retraining.
  • RAG is auditable: Every asset in a generated Parcours can be traced back to a specific catalog item with metadata (curriculum strand, grade level, format, cultural tags). Essential for building trust with teachers and school boards.
  • RAG prevents hallucination by design: The model cannot generate content that doesn't exist in the catalog. It can only retrieve, rank, and sequence existing certified assets.
RAGVector EmbeddingsLLMsCultural Weighting LayerAnonymous Class Code System

Cultural Weighting Layer

A standard RAG pipeline ranks results by semantic relevance. IDÉLLIA adds a re-ranking step that boosts assets meeting TFO’s mandate criteria:

  • Content tagged with Franco-Ontarian regional accents
  • Content featuring Indigenous voices, perspectives, or knowledge systems
  • Content representing a "plural Francophonie" rather than Standard European French

Anonymous Session Architecture

TFO’s privacy policy is non-negotiable: no student email addresses are collected. Standard user-account-based distribution was off the table.

1Teacher generates a Parcours and clicks "Deploy to Class"
2IDÉLLIA bundles selected assets into a temporary session container
3A unique, time-limited Class Code is generated
4Students enter the code at idello.org/code — no account, no email, no tracking
5The session expires after the defined class period

System Architecture

[ Teacher ]
    │
    ▼
[ Natural Language Input ]
"Grade 5 Water Systems lesson, Indigenous perspective"
    │
    ▼
[ Query Parser ]
Extracts: Grade Level · Curriculum Strand · Format Preferences · Cultural Tags
    │
    ▼
[ RAG Retrieval Engine ]
Semantic search across 15,000-asset vector index
    │
    ▼
[ Cultural Weighting Layer ]
Re-ranks results to surface Indigenous / Franco-Ontarian content
    │
    ▼
[ Parcours Assembler ]
Sequences assets into Hook → Instruction → Application arc
    │
    ▼
[ Learning Journey Output ]
Structured lesson with 3–5 certified assets
    │
    ├──► [ Teacher Preview & Edit ]
    │
    └──► [ Class Code Generator ]
              │
              ▼
         [ Anonymous Student Session ]
         idello.org/code → time-limited access, no login

Key Technical Constraints

Vector Index

Built on Idéllo's existing CMS metadata, enriched with semantic embeddings

Retrieval Scope

Strictly bounded to TFO's asset catalog — no external web retrieval

LLM Role

Orchestration and sequencing only — does not generate new content or suggest external resources

Session Data

Ephemeral — Class Codes expire and generate zero persistent student data

Tradeoffs & Prioritization

These are the decisions that shaped IDÉLLIA — not because they were easy, but because they were deliberate.

Tradeoff 1 — Anonymous Access vs. Personalized Accounts

What we chose

Anonymous Class Code model with no student accounts.

What we gave up

Personalization, progress tracking, adaptive sequencing based on individual history.

Why

TFO's privacy mandate made student accounts a non-starter institutionally. Anonymity is the feature that allows IDÉLLIA to be used in schools where other digital tools are banned.

Tradeoff 2 — RAG over Fine-Tuning

What we chose

RAG with a bounded retrieval corpus.

What we gave up

The ability to have the model 'learn' pedagogical preferences over time without re-retrieval.

Why

Fine-tuning would encode the catalog into model weights at a point in time and drift as Idéllo's library grows. RAG keeps the ground truth live and auditable.

Tradeoff 3 — Walled Garden vs. Hybrid Retrieval

What we chose

100% retrieval restricted to TFO's certified catalog.

What we gave up

The breadth and flexibility of open-web retrieval (e.g., linking to Khan Academy, Wikipedia).

Why

This was a values decision as much as a technical one. One unsafe YouTube link erodes the entire trust model. The Walled Garden is not a constraint — it is the product.

Tradeoff 4 — Teacher-Facing MVP vs. Student-Facing Features

What we chose

Build for the teacher first. Student experience in V1 is simply access via Class Code.

What we gave up

Student-facing features like in-session progress indicators, peer activities, or gamified engagement layers.

Why

The teacher is the decision-maker and adoption gateway. If Gabrielle doesn't use IDÉLLIA, her students never see it. A focused teacher experience in V1 was a distribution strategy.

Execution & Rollout

Phased Rollout Strategy

1

Closed Pilot (Pre-September)

A small cohort of 20–30 teachers from French-language school boards in Ontario. The goal was signal quality, not adoption metrics — pedagogical soundness, Class Code reliability under real classroom conditions, and cultural weighting accuracy. Teachers were treated as partners, not users.

2

Board-Level Launch (September)

Timed to align with the school year start — the highest-intent period for teacher planning. School boards were the procurement decision-makers. The launch centred on three deliverables: a demo with real curriculum standards, proof of the privacy compliance model, and a direct “45 minutes vs. 2 minutes” benchmark comparison.

3

Platform Integration & Home-School Bridge

Extending IDÉLLIA’s reach to the “Home Facilitator” persona — parents using “Apprendre à la maison.” This phase added the “Generate Home Link” feature, surfacing gamified activities and Boukili books that reinforced classroom learning at home.

Stakeholder Communication Strategy

School Boards

Lead with privacy compliance and curriculum alignment — before any classroom deployment discussion.

TFO Leadership

Position IDÉLLIA as a platform differentiator, not a feature addition.

Teachers

Lead with speed and cultural authenticity in the first demo — before any conversation about technical architecture.

Parents & Community Orgs

Lead with cultural identity and transparency on the anonymous access model.

Metrics & Impact

North Star Metric

Time-to-Lesson

The time elapsed from a teacher’s first prompt to a deployed Class Code in students’ hands.

Baseline

45+ min

Target

<2 min

145K+

Monthly Active Users

Existing Idéllo platform scale at launch

>95%

Curriculum Alignment

Target accuracy for retrieved assets

0%

Hallucination Rate

Binary quality gate — no external links

Core PM Metrics

MetricBaselineTargetSignal
Time-to-Lesson>45 min<2 minCore value delivery
Parcours Completion Rate>70%Teachers completing full deployment
Class Code Redemption Rate>60%Students actually accessing content
Asset Utilization DepthLow (surface catalog only)Deep catalog accessLong-tail discovery unlocked
Teacher Return Rate (Day 30)>50%Habit formation, not novelty

AI-Specific Quality Metrics

Curriculum Alignment Accuracy

>95%

% of retrieved assets correctly tagged to the specified Ontario curriculum strand. Below this, teachers can't trust the output without manual verification.

Cultural Relevance Score

>90%

% of Parcours (with cultural specifiers) that surface at least one culturally-tagged asset.

Hallucination Rate

0%

Binary quality gate. A single hallucinated external link is a critical incident given the trust model.

Teacher Override Rate

<20%

How often a teacher manually swaps or removes an AI-recommended asset. A high rate signals retrieval quality issues.

Lessons Learned

Lesson 1

The constraint is the product.

Every time the Walled Garden felt limiting — “what if we allowed one external link?” — the answer was the same: the constraint is what makes the product trustworthy enough to use in a school. In regulated or trust-sensitive domains, your constraints aren’t obstacles to creativity — they are the brief.

Lesson 2

Cultural specificity is a product moat.

Generic AI platforms can be faster, bigger, and better-funded. They cannot be more culturally specific. TFO’s mandate to represent a “plural Francophonie” is something no global AI company can replicate without years of localized data collection. Cultural nuance is a durable competitive advantage.

Lesson 3

Anonymity is underrated as a product value.

Personalization gets most of the attention in AI product design. IDÉLLIA taught me that in certain contexts, the absence of tracking is the feature that unlocks adoption. The Class Code model works precisely because it asks nothing of the student. Zero friction. Zero data. Zero barrier to access.

Lesson 4

The teacher is the distribution channel.

Early in the PRD process, I caught myself designing for the student experience. The pivot: in K-12 edtech, the teacher is both the user and the gatekeeper. Building a focused, excellent teacher experience before any student-facing features wasn’t a scope decision — it was a distribution strategy.

Case study by Abdoulaye Binta Bah · Product Manager

Abdoulaye Bah | AI Product Manager Portfolio