[ ← Home ]

Icebox

Ideas I thought about and did not build.

None of these are secret and none of them are coming. If one of them is your kind of problem, take it — I would rather it existed than stayed mine.

Iced 2026
Dopeness

Poker bot for online play

A serious poker agent built on counterfactual regret minimisation, in the line of the Libratus and Pluribus work. A direct application of the CFR material I am working through, and one of the cleaner testbeds for imperfect-information games.

Why notMost sites forbid bots, so it stays a research exercise rather than something to run
Un-iceTreating it purely as a benchmark against open agents
Iced 2026
Dopeness

Autonomous garbage collection

An autonomous driving stack under a robotic arm, collecting garbage without a crew. Smaller machines would take what the truck cannot reach: alleyways, loose litter, the places a route misses.

Why notAn autonomous vehicle stack is a company, not a side project
Un-iceStarting at the small end — one alley, one machine — instead of the truck
Iced 2026
Dopeness

Autonomous park cleanup

Robots that clear litter from national and state parks without supervision. The same platform — navigation, perception, autonomy over unstructured ground — carries over to surveying and aid distribution, which is where it gets interesting beyond cleaning.

Why notHardware, field testing and capital that a student does not have
Un-iceA hardware partner, or a park willing to host a pilot
Iced 2026
Dopeness

Volunteer distributed training network

A voluntary network of machines donating compute to train deep learning models, gathering game data as it goes. Stragglers, verification and bandwidth are the real problem here, not the model.

Why notVolunteer compute is a coordination and trust problem more than a training one
Un-iceA workload where slow, unreliable, mismatched nodes are genuinely fine
Iced 2026
Dopeness

Multi-agent learning for finance

Several learning agents interacting inside a market, rather than one model predicting a series. It depends on the order book work: without a believable simulator there is nothing for the agents to learn against.

Why notNeeds the single-agent groundwork first, and a simulator worth trading in
Un-iceA limit order book simulation solid enough to host agents
Iced 2026
Dopeness

Pre-flop-only poker simulator

Eight players, pre-flop betting, then a random runout to settle the pot. Stripping out every post-flop decision isolates pre-flop strategy, which makes it a clean testbed for regret minimisation before taking on the full game.

Why notSmall enough that it keeps getting passed over for bigger ideas
Un-iceAn afternoon
Iced 2026
Dopeness

Work through an autonomous robotics textbook

Read an autonomous robotics textbook properly and build the projects in it. It is the prerequisite for both cleanup ideas — neither is a software problem I can reason my way through without the fundamentals.

Why notA long read with no deadline attached to it
Un-iceCommitting to either cleanup idea, both of which need the theory first
Iced 2026
Dopeness

AI tutor for the MCAT

A tutor that builds a study plan from a student's own practice performance instead of handing everyone the same schedule. The hard part is not the model — it is showing the plan actually moves scores.

Why notA crowded market where the value sits in content and outcomes, not the model
Un-iceEvidence that a personalised plan beats a good question bank
Iced 2026
Dopeness

Betting lobbies for mini-games

Lobbies where players join and bet on short mini-games, starting with toy games and growing toward larger ones as the format proves itself. The engineering is tractable; the regulation is the actual project.

Why notReal-money wagering is a licensing problem long before it is an engineering one
Un-iceA jurisdiction and a legal structure worth building inside
Iced 2026
Dopeness

Fundamental quantitative finance projects

PCA, linear regression, the one-fund theorem, covariance analysis — implemented rather than read about. These are the pieces every later finance project quietly assumes you already have.

Why notGroundwork rather than a project, so it keeps losing to something flashier
Un-iceWanting a real foundation under the forecasting and multi-agent work
Iced 2026
Dopeness

RNNs for financial forecasting

Recurrent models over financial time series for forecasting and analysis. The honest version of this spends most of its effort on why the naive baseline is so hard to beat.

Why notSequence models on price data mostly rediscover that markets are close to efficient
Un-iceA dataset with real structure in it, or a target other than raw price
Iced 2026
Dopeness

ARC Raiders resource-farming bot

An agent that farms resources in ARC Raiders with nobody at the keyboard. The perception and control problem is the appealing part; the terms of service are why it sits here.

Why notAutomating a live game is against the terms of service of most of them
Un-icePointing it at a game or simulator that welcomes agents