Glasshouse Fund is an autonomous AI that manages a simulated $1,000,000, long-only portfolio. Every trading day it researches, debates, decides, and trades — and publishes all of it: the reasoning, the evidence it cited, the guardrails it ran through, and how its predictions actually turn out. It's a paper fund and an engineering project, not a product. The point isn't returns — it's showing the work.
The AI can be creative; the risk layer is boring on purpose. These limits are enforced deterministically, in code the model cannot argue past — the LLM proposes trades, the guardrails dispose.
Stop-loss and take-profit exits are generated by the risk engine itself and override any LLM trade on the same name — the model can't talk the fund out of cutting a loser.
The AI may only trade names on this watchlist (plus SPY / QQQ for benchmarking). The list leans into the AI-compute build-out — inspired in part by what serious AI-thesis funds are long.
Everything model-facing flows through one gateway, so routing, validation, retries, tracing, and cost tracking are one-time costs — not per-agent ones.
gpt-4o-mini today, behind a provider-agnostic gateway with a cheap/strong tier split (cheap: analysts & summaries; strong: PM synthesis & judges). Swapping in a stronger model or provider is a config change.Full diagrams and the honest retrospective are on the Architecture page.
Nothing is hidden. Read the reasoning, check the scoreboard, browse the code — or point Claude at the fund's read-only MCP server and ask it anything.
Pradnya Wakchaure built Glasshouse Fund as a hands-on platform for modern AI engineering — LLM orchestration, evals, retrieval, agents, and the infrastructure that keeps an autonomous system honest — and as an experiment run publicly, in the open.
Not investment advice. Glasshouse Fund trades simulated capital in a paper account. Nothing here is a recommendation to buy or sell any security. Prices and data may be delayed or imperfect. Past simulated performance says nothing about the future.