The AI infrastructure for social mobility. Breakthrough Social Enterprise · for the Social Tech Trust AI & Social Mobility Challenge Prize 2026.
This is a working prototype of the AI layer described in our prize pitch. It turns the weakest line in any honest application — "the AI layer is in active development" — into something a judge can click through at Parliament.
A technology platform (not a course) that uses AI to do three things for young people shut out of the labour market:
| # | Component | What it does | Prize challenge it answers |
|---|---|---|---|
| ① | Personalisation Engine | Predicts dropout risk from attendance, engagement and disclosed barriers — and recommends the specific intervention to a coach, before someone disengages. | "Support young people disengaged from education, work or training." |
| ② | Employer Translation Layer | Gives associates real-time, accessible feedback on their employability and transferable skills, and reframes lived experience into the language employers trust. Two-sided: also gives employers the context to see capability they'd otherwise screen out. | "Accessible AI tools that give young people real-time feedback on their employability and transferable skills." |
| ③ | Evidence Infrastructure | Turns delivery into live, longitudinal outcome data (6/12/24 months) by cohort, geography, delivery type and population segment — what commissioners and funders see in real time. | Impact embedded in the model itself (the criterion weighted most heavily). |
Built for the hardest cases first (justice-involved), proven across NEET and lower socio-economic cohorts. Scaling to adjacent groups is a reduction in complexity from our baseline — an unusual and genuinely powerful claim.
No build. No server. No internet required.
Just double-click
index.html— it opens in any browser and runs fully offline.
This is deliberate: a Parliament demo should have zero dependencies and nothing to fail on stage. The platform is pure HTML/CSS/JavaScript with no frameworks and no network calls.
If you prefer to serve it (optional, e.g. for a clean URL):
# any static server works, e.g. if you later install Node:
npx serve .- Overview — the one-screen story. Headline impact + the three components.
- Coach Console — the risk board. Note Marcus T. (housing/release) and Sean M. (probation clash) flagged before dropout, each with a specific recommended action.
- Click an associate → the AI explains itself: every risk driver, every recommendation, with a confidence level and a human-in-the-loop disclaimer.
- Employability Studio — switch associates, watch the radar + real-time feedback change. Click Translate for employers to see lived experience reframed into capability language.
- Employer Portal — the other side: candidate context cards funded by corporate D&I budgets.
- Impact Observatory — the live evidence layer: population mix, 24-month retention by circle.
- The Pitch — how this scores against all four judging criteria.
ascent-platform/
├── index.html # entry point — loads everything, zero build
├── assets/
│ └── styles.css # design system (Breakthrough brand: #FFD000, Work Sans)
├── src/
│ ├── data.js # dataset (headline stats verified 8 Jun 2026; records synthetic)
│ ├── ai.js # the AI engine: risk model, recommender, feedback, translator
│ ├── charts.js # hand-built inline-SVG charts (no chart library)
│ ├── views.js # all screens
│ └── app.js # shell, hash router, role lens
└── docs/
├── PITCH.md # the written pitch / seed for the application answers
└── PRIZE-NOTES.md # prize research: criteria, challenge questions, what they value
Two modes, by design:
- Simulated (default). Transparent, deterministic, explainable reasoning that runs offline. Every output exposes the signals it's based on — no black box. This is the right choice for a live demo, and the right value for a justice / social-mobility context.
- Live Claude (optional). Add an Anthropic API key in Settings and the Narrative
Translator upgrades to live
claude-opus-4-8generation. The key is stored only in your browser. This shows the production direction without making the demo depend on the network.
In production, these models train on Breakthrough's 700+ associate longitudinal dataset — spanning prisons, NEET and community cohorts — a dataset no EdTech competitor can access.
- Headline figures (700+ associates, 95%+ engagement, 1% reoffending, 4 employer partners) are verified against the Breakthrough Stats Source of Truth (8 June 2026).
- Individual associate records are synthetic — they model the shape of the real data without exposing any real person. Do not present them as real individuals.
- Cohort-level breakdowns (population split, regional rates) are illustrative in this prototype; production pulls from live delivery records.
The application should describe this honestly: a proven delivery organisation building a venture on an evidence base no startup could replicate, with the AI layer as a working prototype (this) heading to production.
Ascent · The Second Chance OS · Breakthrough Social Enterprise · prototype v0.1 Companies House 12506717 · wearebreakthrough.co.uk