Skip to content

B2C adaptive learning (staged)

Kwilo B2C remembers what a learner knows and doesn't, per topic, across months, and uses it to pick what comes next. Kwi is the voice of that memory. This doc stages the build so each step ships alone and is visibly smarter than the last, instead of the all-or-nothing engine that stalled before (the Study Plan concept-graph tiers sat at 0 rows in prod because everything depended on everything).

Why this exists

  • Who asked: self-initiated strategy call (Bhanu, 2026-07-16). Hypothesis, flagged as such.
  • User pain: a chat tutor alone is commoditized; ChatGPT is free and every kid uses it. A learner practices on Kwilo today and the product forgets every answer. The Revise intent has nothing real to revise from. Parents deciding whether to pay see no evidence the product knows their child.
  • Cost of not doing: B2C retention stays session-shaped (kid churns when homework is done) and there is no parent-facing progress story, which is the willingness-to-pay driver in this market.
  • Validated or guess: the retention mechanics are well-evidenced in the field (mastery loops, spaced repetition effect sizes); that Kwilo's users will feel them is a guess until stage 2 ships and D7 moves.
  • How we know it worked: a learner who got fraction questions wrong on Monday gets a Kwi nudge about fractions on Thursday, does the round, and D7 return for practice users beats the pre-launch baseline. Sean Ellis: turn off topic memory for a week and returning kids notice Kwi got dumber.

The reframe

Do not build "adaptive learning" as one thing. Build the learner model in stages. Stages 2 and 3 deliver most of the perceived adaptivity for a fraction of the full build. The concept graph comes last, only if the cheap stages prove retention.

Stages

Stage 1: capture signal (small, immediately after Kwi Phase 1)

Log per-topic correctness from practice answers that already flow through the product. Today this signal is thrown away.

  • Reuse existing practice-session answer records if they persist per-question results; otherwise add an attempt event on the answer path.
  • Capture time-per-question alongside correctness: time distinguishes "guessed fast" from "struggled to the right answer" at near-zero extra cost.
  • New aggregate: per (user, subject, topic): attempts, correct, avg time, last_seen_at. One table or materialized rollup, recomputed cheaply.
  • No UI. No behavior change. Pure data foundation.

Stage 2: mastery v1 (rule-based, no ML)

Rolling per-topic accuracy + recency = weak/ok/strong per topic.

  • Kwi's recommended action upgrades from goal-based to mastery-based: "You got 3 fraction questions wrong. Today's round is fractions."
  • The Revise intent pulls actual weak topics instead of generic suggestions.
  • Confidence tap as a third signal: after an answer, Kwi asks how sure the learner was (one tap). High-confidence-wrong outranks everything in the review queue; correctness alone can't see it.
  • Cold start: 5-question diagnostic on first practice, framed by Kwi as a game, never as a test. It doubles as placement: prove a topic and skip its basics (pre-assessment skip is the most common activation signal in adaptive systems).

Stage 3: spaced review queue

Wrong answers resurface after 1, 3, and 7 days. Highest proven effect size per unit of engineering in this whole plan.

  • Queue per user; Kwi fronts it: "Remember this one from Tuesday?"
  • Re-serve ONLY the missed questions, never the full set: repeating known material reduces engagement and adds cognitive clutter.
  • Feeds the home recommended action when the queue is non-empty.

Stage 4: difficulty steering

Within a session, next-question difficulty follows rolling accuracy. Needs difficulty metadata on items (PYQ bank has topic tags; difficulty tagging where missing is part of this stage).

Stage 5: concept graph (gated on stages 1 to 3 moving D7)

Prerequisite DAG so the sequencer can say "fractions are weak because division is weak."

  • For grades 1 to 10: LLM-drafted graph per syllabus, human-reviewed. Much cheaper than the exam-corpus ETL that stalled for classes 11 to 12.
  • Unlocks real Study Plan for grades 1 to 10 and honest mastery for 11 to 12.

Parent progress report (the revenue tie-in)

The same aggregates power a parent-facing view: "weak in fractions, improved in geometry, 4 practice days this week." Kid gets Kwi; parent gets evidence. This is the strongest monetization argument for the whole learner-model investment and becomes its own feature doc when stage 2 lands.

Non-goals

  • No BKT/IRT or any ML mastery estimation before stage 5, if ever. Rolling accuracy first.
  • No "AI adaptive learning" marketing claims before stages 1 to 3 are live. BYJU'S burned parent trust on over-claiming; we do not touch that stove.
  • No B2B surfaces in these stages.
  • No cross-subject inference (good at physics says nothing about history).

Metrics

  • North star: D7 return rate of practice-active B2C learners.
  • Secondary: review-queue completion rate (stage 3), repeat practice on weak topics (stage 2).
  • Guardrail: practice session completion rate must not drop when difficulty steering lands (stage 4).

Relationship to Kwi

Kwi (see Kwi persona) is the delivery surface for every stage. Adaptivity without a voice is an invisible algorithm; the model is what makes Kwi feel like it knows the child.