rachel phin

math is forever

email linkedin github
↓

Projects

anki-fox
3.2 Scaled Dot-Product Attention

The input consists of queries and keys of dimension dk, and values of dimension dv. We compute the dot products of the query with all keys, divide each by √dk, and apply a softmax function to obtain the weights on the values.

In practice, we compute the attention function on a set of queries simultaneously, packed together into a matrix Q. The keys and values are also packed into matrices K and V.

✛
⌥⇧A
→
🦊 Anki Fox
New Card
Q: How is attention computed in matrix form?
A: Queries, keys, and values are packed into matrices Q, K, V and attention is computed simultaneously on the full set.
Card
Q: Why divide by √dk?
A: To prevent large dot products from pushing softmax into regions with vanishing gradients.

Anki Fox

Screenshot anything, get Anki flashcards. Captures your screen or YouTube video, sends it to an AI model, and cards appear in Anki automatically. Works with Claude, GPT, Gemini, or a free local model.

Python Multi-model AnkiConnect YouTube
View on GitHub →
pancake.local
Pancake
Active (2)
Ship recurring tasks today
Fix mobile layout ↻ daily
Projects
Semper Paratus
Anki review ↻ daily
Plan next sprint
🎤

Pancake

Priority tracker for households. Shared profiles, task assignment with email alerts, recurring tasks, voice input, and an AI assistant — all backed by plain Markdown files you can edit in Obsidian.

Python Claude API Multi-user Mobile-first
polymarket-trends
Strategy Performance
Win Rate
68%
Avg Return
+12.4%
Trades
847

Polymarket Trends

Real-time trade collector and analytics dashboard for Polymarket. Tracks every trade, surfaces top markets, and backtests strategies against historical data.

Python DuckDB Flask
Claude● connected
Fix the auth bug in pancake
I found the issue — the session token isn't being refreshed after cookie expiry. I'll fix the middleware and add a test.
Ship it
Done. Committed, tests pass, deployed to production.
Message Claude...

Walkie

Claude Code from your phone. Runs on your server inside an OS-level sandbox, accessed through a mobile web UI. Reconnects seamlessly after disconnects with event replay. Your repos, your tools, full autonomy.

Node.js WebSocket Claude Code Mobile
transformer.5.161.182.15.nip.io
Act 2: Architecture
5 Embeddings
6 Attention
7 Multi-Head
Act 4: System Design
20 Review
Knowledge Check
Test your understanding of transformer concepts
Why do we divide by √dk in attention?
A To normalize the output vectors
B To prevent softmax saturation
C To reduce computation cost

Transformer Lab

Interactive learning app for understanding modern LLM architecture. 20 modules covering the journey from RNNs to RLHF, with D3.js visualizations, break-it experiments, and knowledge checks.

SvelteKit D3.js TypeScript GPT-2
github.com/vishnya/LeanAgent
⑂ Fork
LeanAgent
-- Lifelong learning for theorem proving
theorem fermats_little (p : Nat) :
  Prime p → ∀ a,
  a ^ p ≡ a [MOD p] := by
    lean_agent.prove
    -- QED (discovered automatically)

LeanAgent

Lifelong learning framework for formal theorem proving. Continuously generalizes to expanding mathematical knowledge without forgetting previously learned proofs.

Lean 4 ML Research Theorem Proving
ux-checks
✓ touch targets ≥ 44px
✓ no horizontal scroll
✓ font readability ≥ 11px
✓ button uniformity
✓ column alignment
✓ no clipped elements
✓ popups fully visible
7 checks passed (mobile + desktop)

UX Checks

Reusable Playwright test library for catching mobile UX bugs before users do. Touch targets, overflow, font sizes, column alignment, clipped dropdowns, button uniformity — runs on every push.

Python JavaScript Playwright Testing