Inspiration
In September 2025, two private-credit blowups hit within weeks of each other. Tricolor, a subprime auto lender, filed Chapter 7 amid double-pledged collateral allegations, and First Brands, an auto-parts supplier with billions in off-balance-sheet factoring, filed Chapter 11. Lenders lost hundreds of millions. The warning signs were public days to weeks before, but they were scattered across trade press, SEC filings and rumour sites, and in private credit you can't just hit "sell". We wanted to know whether an agent team grounded in the live web could have caught them in time to act, and could prove it without hindsight.
What it does
youFuhnance is an agent team for a private-credit deal team:
- Scout sweeps the live web through You.com for news on every borrower, servicer and obligor behind the fund's positions.
- Verifier merges duplicate stories, rates every source (SEC or court filing > major press > trade > aggregator), runs a corroboration search, and only alerts on confirmed or corroborated claims, each with a quoted receipt.
- Market desk asks "is the market already pricing this?" using change-point detection and regime switching on liquid proxies, labels the action window open / closing / closed, and ranks hedges.
- Covenant desk maps verified news to likely-tripped contract triggers and produces an ordered checklist: draw stop, reservation of rights, title and invoice verification, cash sweep.
- Dashboard replays any day exactly as it was known then, with charts, news pins with receipts, and "Today for the team" discussion prompts.
On replays, it issued a verified alert 3 days before Tricolor's filing (Origin Bank's SEC 8-K) and 23 days before First Brands' (FT on the Deloitte review). In both cases the market proxies only broke 4 and 24 trading sessions later.
How we built it
- You.com Search API: news sweeps, date-bounded replays and corroboration searches, with highlights extraction for clean receipt quotes.
- You.com Contents API: full-text reads of high-severity articles.
- OpenCrawl: the agents coordinate through a board message bus, each with its own identity, publishing events, alerts, actions and briefs. During development we also used OpenCrawl to coordinate parallel coding agents.
- Local models through Ollama: Qwen3-8B as a veto-only judge and nomic-embed-text for event clustering.
- Quant: Bayesian online change-point detection and Markov switching on yfinance data.
- Replay integrity: Wayback Machine snapshots so historical replays only see text that existed that day.
- Stack: Python, SQLite, and a self-contained HTML/SVG dashboard, built with Claude Code including an automated code-review agent.
Challenges we ran into
- Hindsight leaks everywhere. Web pages change after publication, so snippets carried later headlines. Model fits can rewrite the past, and stored verdicts reflected later evidence. We built an automated leak_check that deletes all future data and requires identical outputs, and it caught real leaks we then fixed.
- Small LLMs are too agreeable. An 8B model accepted corroboration that keyword evidence didn't support, so we made the LLM veto-only. Models can reject, never upgrade a verdict.
- Noise: "Tricolor" also means plants, LED lights and flags. Syndication made one story look like 16, and paywalls produced navigation text instead of quotes.
- No public market for private loans. There's no BWIC or loan-price data, so we used liquid public proxies and treated the output as a timing signal, not a price.
Accomplishments that we're proud of
- Verified alerts T-3 and T-23 days before two real bankruptcies, before the market moved, with receipts.
- Provable no-hindsight replays: 24 of 24 leak-check cases pass.
- The system correctly held an unverified rumour ("bank seizes Tricolor", aggregator sites only) at watch, and alerted only once the SEC filing confirmed exposure.
- An action layer that speaks a lender's language: draw stops, title custody, double-factoring checks, before the automatic stay.
- Cheap to run: about $0.60 per full replay of one case in You.com credits.
What we learned
- Live-web grounding is necessary but not sufficient. Source quality and corroboration are what make an alert trustworthy enough to act on.
- In private credit the value isn't "predicting default". It's buying time while the exit window is still open.
- Proving that replays use no hindsight is harder than building the agents, and it's what makes the results believable.
What's next for youFuhnance
- Exit-liquidity desk: dealer appetite and distressed-buyer signals from news-reported prices.
- Live mode: scheduled sweeps on a real watchlist, delivered as a 7am brief to email or Slack.
- NAV loans: monitor PE sponsors and their portfolio companies.
- Ingest real deal documents, so covenant terms come from the credit agreement rather than mock data.
- A larger backtest across more 2023–2025 credit events, to measure precision and false alarms.
