If you want to check my claims rather than take my word for it: the board is live at seiche.info, updates twice a day. Scroll to the bottom of today's letter, the section called "The record, nothing hidden." That's the backtest with the misses listed by name, not just the wins. All the data is free and keyless (FRED, NY Fed, OFR, Treasury, CFTC, ECB) so you can rebuild any number yourself if you're feeling suspicious. Honestly the most interesting thing to poke at is the ML calibration failure. The model ranks risk fine but its probabilities were junk, and the terminal says so on the page. Ask me anything, including the awkward stuff. I'd rather lose the grant than dodge a good question.
1
Seiche: a tamper evident early warning system for dollar funding stress
ProposalGrant
Closes August 21st, 2026$0raised
$5,000minimum funding
$25,000funding goal
Offer to donate
42 daysleft to contribute
You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
Project summary
Seiche (https://seiche.info) is a free public terminal that monitors stress building in the dollar funding system, the repo and money markets that seized in September 2019 and March 2020. It fuses free public data (FRED, the New York Fed, the OFR, Treasury, the CFTC, the ECB) into a forward looking read published twice a day.
The part this community will care about is the epistemics. Every daily forecast is written into a hash chained, as published ledger that ships with the site, so nobody, including me, can rewrite a bad month. Every engine validates walk forward against a naive benchmark and prints the verdict even when it loses. When my ML model turned out to rank well (out of sample AUROC 0.826) but produce miscalibrated probabilities, the terminal published that finding about itself and demoted the model to a ranking tool. Today's headline probability is wrapped in a Venn Abers calibrated interval with finite sample validity guarantees: not "our probability is 7 percent" but "the calibrated probability is provably between these bounds."
Evidence so far: replayed point in time to September 12, 2019, the board reads EROSION and flags September 16, the exact day the repo market broke, as a crunch window. In an orthogonality test that removes the target's own variable family from the signal, the index still catches 69 percent of historical stress events (Wilson 95 percent CI 42 to 87), and the structural components alone sat at the 98th to 100th percentile 42 days before both the September and December 2025 squeezes. The misses are listed by episode, next to the hits.
What are this project's goals? How will you achieve them?
Goal 1: a full year of sealed, as published daily forecasts that anyone can audit, and a public verdict on whether the instrument earned its keep. The ledger is already accruing daily on an automated schedule; the goal is to keep it unbroken and get it independently checked.
Goal 2: an independent methods review by two quantitative researchers, published unedited whatever it says.
Goal 3: a replication package and open API so researchers can rebuild every number from public data.
Goal 4: open source the codebase under AGPL 3.0.
How will this funding be used?
Minimum ($5,000): keeps me on Seiche roughly half time for six months, funds the infrastructure (the whole stack runs on a 36 euro a month server), and ships the replication package.
Full goal ($25,000): converts me to full time for a year, adds the independent methods review ($3,000), and funds documentation and researcher outreach. The terminal, the daily letter and the sealed record stay free either way.
Who is on your team? What's your track record on similar projects?
Solo. I am a lawyer from Jaipur, India who taught herself to code. I came to money markets from below: I built LiquiLens, an early warning system for Indian small business lenders, which backtested on 36 institutions caught nine of ten collapses with a median lead of fifteen months, and this month scored a live microlender below investment grade one quarter before the rating agency. Seiche is the same method pointed at the headwaters those lenders depend on. It went from first commit to a live, 22 engine terminal with roughly 200 automated tests (including one that mechanically forbids look ahead) in about two months, solo.
What are the most likely causes and outcomes if this project fails?
The most likely failure is that the signal does not survive its own forward test: the sealed record accumulates and the instrument turns out to be no better than climatology out of sample. If that happens, the record will show it and I will publish that verdict, which I would consider a useful public result, just not the one I want. The second failure mode is solo founder risk: I run out of runway and the cadence degrades, though the automated pipeline would keep the ledger accruing regardless. The third is attention failure: the instrument works and nobody with decision power looks at it. This grant mostly buys down failure modes two and three.
How much money have you raised in the last 12 months, and from where?
Zero external funding so far; entirely self funded. Two applications are currently pending and undecided: an Emergent Ventures application (submitted July 2026) and an inquiry to Coefficient Giving, both disclosed here for transparency. The site accepts crypto donations, which have so far covered approximately nothing.