SEAP, spoken English practice for young people entering work

Chapter 1 · Who it is for

For students at vocational institutes, spoken English decides who gets hired, and nobody listens to them practice

When about 3,970 learners joined the pilot, we asked them why. Three in four, 73.6%, said placements and job interviews. Two thirds rated their own speaking at the lowest two of four levels.

The learners are students at ITIs, polytechnics and government and lower-tier degree colleges, aged 18 to 35, most from regional-language backgrounds. Their courses teach English on paper. Instructors have no time to listen to each student, and institutes have no standard way to measure speaking.

Two of these rows matter most for what came next. Most had studied in English and still rated their speaking low, so more reading was unlikely to fix it. And a quarter had already tried an English app and stopped, so a product that only scored them would lose them too.

I narrowed the product to this group on purpose

The read-aloud version served any student taking an English course on a national online platform. I narrowed the target to vocational and college learners, because for them spoken English is a hiring filter. The outcome I set was job-linked speaking ability and confidence, measured in the learner, not syllabus coverage.

Program context, kept apart from results

Sources: SEAP 1.0 baseline survey (n about 3,970, May 2026); SEAP decks, March and July 2026; concept notes, October 2025 to January 2026.

Chapter 2 · Three versions

Each version taught us one thing, and the pilot data turned an assessment into a tutor

SEAP started as a read-aloud check inside online courses, before I joined. I planned two more versions, each to answer a question the one before could not.

1.0 tested whether learners would use it, and its data set 2.0's shape

In November 2025 we split the work into two tracks: 1.0 on the existing speech model, deployed through an implementation partner, while 2.0 and a better model were built. That bought months of real use, at the cost of a model not tuned to Indian accents and a team carrying two products.

I wrote the pilot to answer operational questions: can institutes onboard learners, do learners activate, can we track completion. The answer pulled two ways: 44% retried unprompted, so learners wanted practice, but 6 in 10 never came back, so they also needed a path.

User research set 2.0's scope

I led user research with about 100 vocational and college students in small towns. More than half had already tried Duolingo, and most had stopped. With the designer I wrote the research guide, interview questions and vignettes, ran workshops to fix scope, visited learners after 1.0 went live, and tested 2.0 prototypes with them.

Sources: SEAP 2.0 note, November 2025; SEAP concept notes; SEAP decks, March and July 2026; my handover notes, July 2026.

Chapter 3 · Pilot design

I set the pilot's bar at a return visit, so a supervised class could not fake success

A supervised session in a computer lab can produce hundreds of registrations in an hour. It cannot produce a learner who comes back the next day on their own. So I counted a learner as Reached only after a return.

I wrote the pilot design on 27 April 2026, before the first institute went live. Reached meant three sessions, with at least one a day or more after the first. Registrations were reported but never contracted. For 2.0, part of the implementation partners' fees was tied to Reached, so the partners were paid for learners who came back.

6 in 10 never came back, so 2.0 needed a path

That is 39% of the learners who took part, so 6 in 10 never came back to count. Learners given open practice, with passages in random order, did not return. That is the problem 2.0's ordered path was built for, and it fixed it. A committed group practiced hard: the top quarter of users did about 60 assessments each over the eight weeks. The average hides both groups.

Field logs surfaced a login problem and a metric bug

Of nine issues logged in the field, five were learners not receiving the one-time password to log in: the cost of going standalone. Another mattered more for measurement. The dashboard showed zero activated for learners who had returned more than a day later, and we traced it to duplicate names. A metric that reads zero while learners practice teaches everyone to ignore it.

Sources: SEAP 1.0 pilot design, 27 Apr 2026; field issues and user feedback log, May to June 2026; my handover facts sheet, July 2026; SEAP deck, June 2026 (top-quarter usage).

Chapter 4 · The signal

Learners retried passages when they could see a score rise, and said their real problem was making sentences

Without being asked, 44% of learners read a passage they had already done a second time. On the retry, their median fluency rose 7 words a minute, from 87.6 to 94.6.

The gain is on a passage learners had already read. What mattered was the behavior: it showed what would make learners practice, seeing themselves improve.

What learners found hard set 2.0's focus

The baseline survey pointed the same way from the other side.

78% of survey respondents said they felt more confident

In the end-of-pilot survey, 78% of 537 learners who chose to answer said they felt more confident speaking English. Self-report, 537

The July deck cites the retries as evidence for centering 2.0 on practice. The survey pointed to what to practice: making sentences, more than reading.

Sources: SEAP deck, July 2026 (retries, median speed, end-of-pilot survey); SEAP 1.0 baseline survey, May 2026.

Chapter 5 · Learning design

2.0 is a 25-lesson path where speaking and listening lead, and grammar and vocabulary hold them up

In 1.0 a learner picked a passage and read it aloud. In 2.0 she follows 25 lessons in order, from the basics to being ready for an interview, set in workplace and interview scenes.

Speaking stays at the center, because speaking is what employers screen for. Grammar and vocabulary come in to support it, because those were the gaps learners named.

Levels are mapped to the CEFR, and 70% of learners moved up one level

Most pilot learners were beginners. With the pedagogy team I defined levels mapped to the beginner bands of the CEFR, an international scale of language ability. Moving up means finishing the 25 lessons above a threshold score. At scale, 70% of learners moved up one English proficiency level.

A diagnostic test at entry places each learner at the right level.

The habit loop was designed around a web app's weaker notifications

A native app sends reliable notifications, the best tool for bringing learners back. 2.0 is a progressive web app, so the loop leans on other things.

Instructions are translated into Telugu and Kannada to scaffold beginners. The English being learned stays in English.

One question was still open at handover: a leaderboard scored by points earned rather than by percentage correct, so that a retry earns points instead of dragging a score down. Open at handover Chapter 8 sets out the test.

Sources: SEAP deck, July 2026; SEAP 2.0 priorities and design decisions in my handover notes, July 2026; baseline survey, May 2026.

Chapter 6 · Decisions

Each 2.0 decision had a cost, from harder impact measurement to weaker notifications

The product, pedagogy and design leads and I decided scope and flows together, with ML and engineering joining for technical calls, and I had the final say. We ranked features by reach, impact, confidence and effort. Exercise types were kept to the smallest set an instructor could understand in one sitting. The larger calls are below.

The speech data was run like a product, with its own quality loop

I designed the flow of a custom collection app (login, consent, an audio check, district selection), which another engineering team built. Vendors collected audio across India's roughly 800 districts, with at least 40% of each district's data planned from each gender. Each collector's recordings were checked in rounds of ten. If four of the ten failed, all of that collector's recordings could be rejected, and a district short of its quota for one gender was sent back for re-collection. Two-way conversation practice was parked for lack of model capacity and budget.

Measured The result is one of India's largest speech data programs: more than 5,000 hours across about 800 districts. The speech model fine-tuned on it is 50% more accurate. The pilot's scores came from the base model; building the new model into the product is the next step (chapter 8).

Sources: My handover notes, July 2026; data collection and annotation RFPs, October and November 2025; quality control plan and SOP; recollection rules in the quality log; SEAP deck, July 2026.

Chapter 7 · Measurement

I wrote the measurement plan before the scale decision, with pass marks set in advance

One rule ran through every SEAP number: each carries its source and its filter, and a target is never shown as a result. This page follows the same rule.

The plan was built to be able to say no

In May 2026 I wrote the pilot's metrics framework for the July decision on whether to scale 2.0. Its job was to make a no possible: five pass-or-fail gates, fed by seven critical-path metrics, each with a pass mark written down before the data came in. Behind them sat a register of 146 metric definitions, so no two teams reported the same number two ways.

The decision in July was to go ahead: 2.0 began deploying on 23 July 2026 and scaled through state partnerships to 1M learners.

For 2.0 I set the next priorities in order: certificates generated automatically, WhatsApp nudges to push completion, and short in-app questions after lessons 10 and 25 to measure confidence inside the product. A triage framework, with thresholds we could adjust, ranked institutes to direct field staff's effort.

Chapter 8 sets out how I would measure 2.0.

Sources: SEAP 1.0 metrics framework, May 2026; 2.0 priorities in my handover notes, July 2026.

Chapter 8 · What comes next

Each result has a next test, and 2.0's starts with the leaderboard

The pilot and 2.0 gave measured results and two survey results. Each one points to a next step, under the same rule: the measure and its pass mark are written down before the data comes in.

The first test on 2.0 is the leaderboard

The first experiment I would run answers the leaderboard question left open at handover: rank learners by points earned or by percent correct. Classmates compare ranks, so I would assign whole institutes to one version, with Reached within 14 days of joining as the primary metric and first-attempt scores on unseen lessons as the guardrail.

Then build free conversation practice

After that, I would build free conversation practice once the speech models allow it.

Sources: Chapters 3 to 7; SEAP 1.0 metrics framework, May 2026; 2.0 priorities and the open leaderboard question in my handover notes, July 2026.

Reflections

The return-visit bar and the baseline survey together shaped 2.0