Finance runs on trust, and most AI can't earn it yet. Money Talks 2 is one day to change that: build AI that finance teams would actually use, then prove it works.

Money Talks 2 is an AI x Fintech hackathon hosted by Block Convey and localhost:nyc, with Bond AI. Engineers, founders, designers, and finance operators get one day in Brooklyn to build real AI products for finance.

Here's what makes it different: you don't just build, you prove it. Every team connects its agent to PRISM by Block Convey to watch it run, find where it fails, fix it, and show the score that proves the fix. Every team also builds on NicrronNinoo, which gives you one API for 80+ AI models plus an AI app builder to ship your demo UI fast.

We're looking for: working products, agents, tools, and workflows that solve a real problem in finance, backed by evidence that they work.

We're not looking for: slide decks, mockups with nothing behind them, or projects built before today.

Read this first: the three tracks are themes, not assignments. Pick the one that interests you and build anything inside it, whether that's an agent, a tool, a full product, or a workflow. It's wide open. The ideas below are starting points for when you're stuck, not a menu you have to pick from.

Only two things are required, whatever you build:

1. Build on PRISM by Block Convey and show the three PRISM proofs for your track (listed below).
2. Build with NicrronNinoo: your model calls go through the Nicrron API, or your app is built with Ninoo.

Do those two things and the rest is yours. Build something you'd want to exist.

Track 1: Money Intelligence

AI that helps people understand financial information and make better decisions. Think investing, research, analysis, portfolio intelligence, and personal finance.

Starting points: an agent that explains an earnings report in plain English with the key numbers pulled out, a research copilot over filings, or a portfolio Q&A assistant.

Required PRISM proofs:

1. Knowledge Base: upload your filings, statements or research (.pdf, .txt or .md), assign them to your agent, and search them from your agent (kb_search, see docs → Knowledge Base). Show a run whose retrieval step used them.
2. Root Cause: first analyse 5 to 10 failing runs (Traces → tick them → Analyse selected, 1 credit each). Then click Analyze now on the Root Cause page (5 credits). Show a cluster of 3+ failed runs and its cause, plus one sentence from you on why it happens.
3. Evaluator: write your own evaluator in the Evaluators Hub (for example, "every number in {output} appears in {context}", where {output} is your agent's answer and {context} is what it retrieved). Show it scoring v1 and v2 on the same inputs.

Demo: the failure, the fix, and the score that proves it.

Track 2: Money Operations

AI that improves how money moves and how finance teams operate. Think payments, expenses, reconciliation, invoices, CFO workflows, and automation.

Starting points: a reconciliation agent, an invoice tracker that chases late payments, a receipt-to-spending-table pipeline, or a month-end close assistant.

Required PRISM proofs:

1. Sessions: open one run on the Steps tab and show every tool call your agent makes (6+ steps).
2. Alert: set an alert on something that matters to your workflow (latency, error rate, satisfaction score) and show it firing on a real run. A failed tool call that leaves the run with no answer counts toward error rate.
3. Root Cause: first analyse 5 to 10 failing runs (Traces → tick them → Analyse selected, 1 credit each). Then click Analyze now on the Root Cause page (5 credits). Show a cluster of 3+ failed runs and its cause, then show v2 on the same inputs no longer failing.

Demo: the failure, the fix, and the alert that would have caught it.

Track 3: Trust & Risk

AI that helps financial systems make safer decisions. Think fraud, risk, compliance, security, and trustworthy financial AI.

Starting points: an agent that flags unusual transactions and explains why, a compliance checker for financial advice, or a chat assistant for loan terms with hard guardrails.

Required PRISM proofs:

1. Guardrail: write your own guardrail for a finance risk (account or card numbers, SSNs, "guaranteed returns", unlicensed advice) in Block mode. Call it before your model (POST /api/guardrails/evaluate with a trace_id, see docs → Guardrails, or route your model through the PRISM proxy). Show it blocking a real run.
2. Block check: send that run's trace under the same trace_id. Then, on the Guardrails page, show whether the block was honoured, not honoured or unverifiable, and tell us if PRISM got it right.
3. Root Cause: click Analyze now on the Root Cause page (5 credits) and show a cluster of 3+ failed or blocked runs and its cause. If none appears, analyse 5 to 10 risky runs first (Traces → Analyse selected, 1 credit each) and run it again.

Demo: the risk, the guardrail, and the run it stopped.

Each criterion is scored out of 5.

• PRISM Proofs (out of 5): Did you show all three required proofs for your track? Judges want real evidence: a failure you found, the fix you made, and the score, alert or block that proves it worked. This is the core of Money Talks.
• Technical Execution (out of 5): Does it actually work, live? Judges look at how reliable the agent is, how well it uses NicrronNinoo and PRISM, and whether the build holds up beyond the happy path.
• Uniqueness of Idea (out of 5): Is this a real problem in finance, approached in a fresh way? Judges reward ideas a finance team or user would want tomorrow over another generic chatbot.
• Design (out of 5): Is it clear and usable? Finance users need to trust what they see, so judges look for clean flows and outputs a person can actually act on.

• PRISM Proofs (out of 5)
  Did you show all three required proofs for your track? Judges want real evidence: a failure you found, the fix you made, and the score, alert or block that proves it worked. This is the core of Money Talks.
• Technical Execution (out of 5)
  Does it actually work, live? Judges look at how reliable the agent is, how well it uses NicrronNinoo and PRISM, and whether the build holds up beyond the happy path.
• Uniqueness of Idea (out of 5)
  Is this a real problem in finance, approached in a fresh way? Judges reward ideas a finance team or user would want tomorrow over another generic chatbot.
• Design (out of 5)
  Is it clear and usable? Finance users need to trust what they see, so judges look for clean flows and outputs a person can actually act on.