Claude helped with these descriptions, but the heart and vision is truly from me and my initial prompt to Claude describing the project and my wishes for where it would go! Hope to present.
Inspiration Hedge funds like Two Sigma and Renaissance Technologies have delivered returns that dwarf public market indices for decades — but their minimum investment requirements lock out everyone except institutional capital. Meanwhile, platforms like Robinhood democratized stock trading but stopped short of giving everyday investors access to the strategies behind institutional performance. Fund LLC is our attempt to close that gap: a pooled capital platform where verified investors open their strategies to the public, and visitors — from individual savers to college endowments — can allocate into funds they'd otherwise never qualify for.
What it does Fund LLC is a two-sided investment platform with three "worlds" — Day/Day, Month/Month, and Year/Year — representing different time horizons and risk profiles. Investors publish a "bar": a public profile showing their verified performance across all three worlds, their fee structure, and their strategy. Visitors describe what they're looking for in plain English, and an AI agent matches them to the best-fit investors with reasoning. On the investor side, a second AI agent parses a natural language description of your strategy and automatically generates your full bar — performance numbers, fee, tagline, risk level, and strategy summary. Fund memberships are stored in Supabase so every visitor can track a portfolio of funds they've joined, with live net gain calculations per allocation.
How we built it The frontend is Next.js 14 with TypeScript, styled with a custom design system built from scratch — DM Serif Display for headings, Syne for UI, DM Mono for numbers. Authentication and the database are powered by Supabase, with Row Level Security policies so visitors only see their own portfolios and investors only see their own clients. The AI agents are built on Claude (Anthropic), called from Next.js API routes — one agent for visitor matching, one for investor profile parsing, both returning structured JSON. We used ClickHouse as our analytics layer and integrated Datadog's Lapdog for local LLM observability, so every Claude call during development shows up as a traced span with token counts, latency, and the full prompt/response. The whole thing deploys to Vercel.
Challenges we ran into Getting Supabase's TypeScript types to play nicely with the generated database client took longer than expected — the generic inference kept collapsing table types to never[], requiring as any casts as a short-term fix. Lapdog's dd-trace initialization was blocking Vercel's build process because it tried to connect to localhost:8126 at compile time, which doesn't exist in a cloud build environment — we fixed this by making the tracer lazy-loaded and gated behind a LAPDOG_ENABLED environment variable. Balancing the three-world paradigm in the data model also required careful thought: a visitor might want to join the same investor across multiple worlds, so fund memberships needed a composite unique constraint on (visitor_id, investor_id, world).
Accomplishments that we're proud of The AI agent UX replacing traditional form inputs is something we're genuinely proud of. Rather than asking visitors to fill out dropdowns for risk profile, cap preference, and world — you just describe what you want in a sentence, and Claude does the rest. Same on the investor side: instead of a form with number inputs for each performance metric, you describe your strategy and the agent populates your entire bar. The design system also came out cohesive — every screen from landing to fund confirmation feels like the same product, built in a single day.
What we learned Lapdog is a genuinely great tool for hackathon-speed AI development — being able to see every prompt, token count, and latency in a browser dashboard without setting up a full Datadog account made debugging the Claude agents fast. We also learned that natural language as a UI primitive has real tradeoffs: it's faster and more expressive for users who know what they want, but requires more careful prompt engineering to get reliably structured output. The as any TypeScript escape hatch is useful under time pressure but leaves real technical debt worth cleaning up.
What's next for fundllc The most important missing piece is real financial data integration — right now performance stats are input manually, and the platform needs a live feed to be credible. Next steps include connecting to a market data API so investor performance updates automatically, building out the investor dashboard with a proper client management view, adding a notification layer (agent-triggered alerts when a fund's performance changes significantly), and wiring up the actual transaction layer so fund inclusion isn't just recorded in a database but executes real portfolio allocation. Longer term, the B2B angle — licensing the platform to Robinhood or similar — requires a compliance and verification layer for investor track records, which is a non-trivial regulatory problem worth solving.
Built With
- supabase
- typescript
Log in or sign up for Devpost to join the conversation.