ShowYourWork Math
A math homework tutor that does not just give the answer but shows its confidence at every step, flags when it is unsure, and only spends extra compute double-checking the steps it actually struggles on.
high school and college students cramming for STEM homework and exams
- Step-by-step solution with a per-step confidence meter ('AI is 92% sure here, 41% sure there')
- Auto-double-check mode that re-solves weak steps before showing the answer, surfaced as 'I wasn't sure, so I checked twice'
- Plain-language explanation of what went wrong when the tutor catches its own mistake
- Side-by-side comparison of two candidate solution paths so students can pick the clearer one
DeepLook just demonstrated a monitor-and-intervene decoding loop that improves math accuracy while cutting token cost by 87%, making confidence-aware tutoring finally cheap enough to give away.
Math homework help is a proven massive market: Photomath has 100M+ downloads and Mathway is a household name, with multiple 'best AI math apps 2025' roundups confirming sustained student demand.Best Math AI Apps: Top Picks for 2025 - AI Tutor Blog ↗Best AI Tools for Math in 2025 ↗
Market is crowded with well-funded incumbents (Photomath, Mathway, Khanmigo, MathPad, Gradily, Vega AI) all offering step-by-step AI math tutoring; the confidence-aware angle is incremental, not category-defining.PhotoMath vs Mathway vs MathPad: Which Math Solver is Best? (2025) ↗Best AI Math Solvers Compared: Photomath vs Mathway vs Gradily ↗
Proven willingness to pay exists (Mathway $9.99/mo or $39.99/yr; Khanmigo $4/mo), but price-sensitive students and free alternatives (Photomath free tier, Khanmigo free for teachers) compress ARPU; the 'give it away' premise conflicts with reality.Mathway Pricing and Packages For 2025 ↗Khanmigo pricing: Free for teachers, $4/month for parents & learners ↗
The DeepLook paper (arXiv 2607.22602, June 2026) is real and confirms the 87% token reduction claim, but selective-confidence decoding is a generic technique that frontier model providers (OpenAI, Google, Anthropic) will bake into base APIs within 1-2 model cycles, eroding any moat.[2607.22602] DeepLook: Deeper Thinking with Lookahead - arXiv.org ↗
Wrapping an off-the-shelf LLM with confidence flagging and selective re-verification is straightforward engineering; only the per-step confidence calibration and confidence-aware compute routing require nontrivial prompt/eval work.MathTutor Pro Pricing | 300 AI Tutor Minutes ↗