PulsarFi

PulsarFi is a 1-to-1 backed stock tokenization platform for Indonesian IDX equities, built on Arbitrum Sepolia.

The idea is simple: Indonesian stocks like BUMIP, ENRGP, and TLKM are illiquid outside trading hours, locked to brokerage accounts, and inaccessible to most DeFi tooling. PulsarFi turns each stock position into an ERC-20 receipt token called a Pulsar Stock, paired against IDRX liquidity pools via Uniswap V4, with a custom hook enforcing the protocol fee at the pool level. Users can trade these tokens 24/7 without leaving their wallet.

Every token in circulation is backed 1-to-1 by a real stock position held by licensed custodians. To mint new supply, a proposal must pass a 3-of-5 on-chain multisig approval across independent custodians before anything gets created. No single custodian can mint alone, and since custodians are competing firms in practice, collusion is economically irrational.

On top of the tokenization layer, PulsarFi runs an autonomous AI trading agent — three cooperating roles (Quasar the supervisor, Nova the market analyst, Comet the executor) built on a CloudWeGo Eino graph. A user describes what they want in plain language; the agent resolves the trade, verifies tradeability and market conditions, and executes on-chain within a budget/expiry/cooldown the user explicitly grants — with every reasoning step hash-chained on-chain for independent audit.

The flow for users is straightforward: connect a wallet, swap IDRX for any listed Pulsar Stock (directly or via the agent), and hold or trade freely. KYC is only required at the redemption step, not for trading. Users who want to exit back to the real stock submit a redeem request, which custodians then process off-chain.

On the custodian side, there is a full management console covering mint proposals, approval queues, redemption processing, KYC access control, and reserve records. Authentication across the whole platform uses SIWE (EIP-4361), so wallet signatures are the only credential needed.

We entered the buildathon with the tokenization protocol already live on Arbitrum Sepolia, and used the building period to ship the AI trading agent end-to-end and harden it against real bugs found through live testing.

On September 13, we shipped the full agent trade execution pipeline: dynamic, fully multilingual confirmation cards (no hardcoded UI strings — every label is generated by the LLM in the user’s own language), and on-chain subtask telemetry so every reasoning step in a trade is independently auditable.

September 14-16 went into a significant architecture rework of the agent’s routing graph — moving to a clean interrupt/resume model where a Task commits the instant intent is recognized, and Nova’s market analysis became a mandatory advisory step before execution rather than optional. Alongside that, live testing surfaced and fixed a cluster of real bugs: a missing sell-side allowance token address, a chat panel fetching a chat row before it existed, and the Arm panel not re-approving allowances unconditionally on every arm.

On September 17, we shipped the scalp-trading routing refactor and redesigned how the agent’s live sub-task state reaches the frontend: instead of the client inferring retry/failure state from incomplete events, the backend now owns the full sub-task list and pushes its complete, current state on every change — the frontend only ever renders what it’s given. We also published the design doc for the next iteration: automated take-profit/stop-loss and position monitoring for swing trades.

Not currently fundraising. The project was built independently for this hackathon with no external investment.