AI Projects

Portfolio of four AI/ML algorithms, each implemented from scratch and running live in the browser — no backend, no ML/algorithm libraries.

Live site: https://kmb-ai-projects.netlify.app/

I originally wrote these as standalone Python scripts (for Columbia CS coursework). I later rewrote every project from scratch in JavaScript so I could give them a frontend and ship them as an interactive website instead of CLI scripts.

Stack

Static site, no build step, no package manager, no framework.

• Each page (2048.html, npuzzle.html, sudoku.html, transformer.html) is a single self-contained HTML file with the algorithm implemented in an inline <script>.
• style.css is shared across all pages.
• index.html is the landing page linking out to each demo.

Run locally with any static file server, e.g.:

python3 -m http.server

Projects

2048 — Expectiminimax Agent (2048.html)

Adversarial search agent that plays 2048 against the game's random tile spawner.

• Expectiminimax search — chance nodes for the random tile placement (90% chance of 2, 10% chance of 4), max nodes for the agent's move choice.
• Iterative deepening — searches progressively deeper until a time budget is exhausted, so the agent always returns a move even under time pressure.
• Alpha-beta pruning on the max/min branches to cut the search space.
• 6-component board evaluation heuristic: empty cell count, corner placement of the max tile, monotonicity, merge potential, max tile value, and smoothness.
• Two search modes toggle in the UI: probabilistic expectimax (true chance nodes, configurable 2/4 weights) vs. adversarial minimax (worst-case opponent).

N-Puzzle Solver (npuzzle.html)

Solves the sliding-tile N-puzzle (8-puzzle / 15-puzzle) with a choice of six search strategies, visualized in real time:

• A* and Greedy Best-First (Manhattan distance heuristic)
• BFS, DFS, IDS (iterative deepening search), UCS
• Solvability check before search (parity of inversions) so unsolvable boards fail fast instead of exhausting the search space.

Sudoku — CSP Solver (sudoku.html)

Sudoku as a constraint satisfaction problem:

• Backtracking search over cell assignments.
• MRV (minimum remaining values) heuristic for variable ordering — picks the most-constrained empty cell first.
• Forward checking to prune domains of neighboring cells after each assignment, detecting dead ends early.
• Step-by-step solver trace with backtrack count, playable live against a bundled puzzle set.

Transformer (transformer.html)

Encoder-decoder Transformer trained live in-browser on a toy sequence-copy task, with backpropagation implemented by hand (no autograd).

• Multi-head self-attention and cross-attention, positional encoding, encoder/decoder blocks.
• Manual forward and backward passes, Adam optimizer, cross-entropy loss.
• Toggleable Pre-LN vs. Post-LN residual/normalization variant and a dropout control, so architectural choices can be compared live during training.

Repo layout

index.html        landing page + nav
style.css          shared styles (theme vars, fonts: Orbitron / Rajdhani / JetBrains Mono)
2048.html          2048 expectiminimax agent
npuzzle.html        N-puzzle solver (A*, BFS, DFS, IDS, UCS, Greedy)
sudoku.html          Sudoku CSP solver
transformer.html    in-browser Transformer training

Each page is independent — there's no shared JS module system, so any logic common to multiple pages is duplicated per file rather than imported.