# LiquiLens

> LiquiLens is an early warning system for bank and lender failure. It reads what institutions and regulators publish, funding structure, filings and market signals, and flags the ones drifting toward distress years before ratings move. Validated in the open on three markets: 16 of 18 non fraud failures caught at a median lead of 21 months across two decades of Indian collapses, 73% recall across 552 US bank failures, and seven named European case files from Northern Rock to Credit Suisse replayed through the same unrecalibrated lenses. Every call lands on a tamper evident, timestamped record. Supervisory dashboards publish raw filings with no synthesis and terminals that synthesize cost tens of thousands a year; LiquiLens is the free middle shelf that states a view and grades it in public, one layer of the same lab as Seiche (funding plumbing, https://seiche.info) and Undertow (market liquidity, https://liquilens-undertow.com). The same engine carries treasury, liquidity risk and credit scoring for lenders and businesses, with live books over India's regulated lenders and every FDIC insured US bank.

## What it does

- Failure Radar: a live board over India's regulated lenders (banks, small finance banks, co-operatives, NBFCs, MFIs, HFCs) reading each institution four ways at once: a failure probability fitted on the real collapses in the record, distance to the RBI's own PCA and SAF action zones, funding fragility from the liability side, and a market implied distance to default for listed names, repriced daily
- Attested record: a hash chained point in time ledger of every watchlist and alert as published, signed with Ed25519 and anchored via OpenTimestamps, so the track record is independently verifiable
- Multi horizon cash and liquidity forecasting for businesses and lenders
- Liquidity at Risk: Monte Carlo simulation of cash shortfalls over a rolling horizon
- Liquidity Credit Score: an LCR and NSFR driven score of buffer adequacy, funding stability and concentration
- Contagion Radar: predicts which counterparties will delay payment before it lands
- Liquidity Mesh: matches surplus cash to deficits and prices it inside the bank spread
- Early Warning: a five signal lender suite (CE trajectory, network contagion, loan stacking velocity, coordinated fraud rings, single party concentration) under one false alert budget, aligned with RBI expectations on early warning systems and expected credit loss
- Environment layer: coverage honest external factor observatories (monsoon and ENSO, food and fuel prices, funding conditions, political stress) for the regions the lenders operate in, refreshed every six hours from free public sources; a dark feed returns an explicit 503, never a calm looking zero (GET https://api.liquilens.in/api/environment/india, /egypt, /world)
- Scoring gate: environment readings are context only and never move a risk score; whether they ever may is decided by pre registered studies on the crisis record, published win or lose. The monsoon study ran in July 2026 with two decades of sourced IMD rainfall data and came back short of the bar, so the gate is CLOSED; the verdict, data provenance and re run trigger are served at GET https://api.liquilens.in/api/environment/gate
- Public signals layer (fifteen layers live as of 20 July 2026): a standalone early warning mode built from free public data alone, no lender book and no supervisory feed. Market regime over the bank index and macro tape; per institution news watch; bond market and allocator watches; Stablecoin Rails Watch (issuer level peg, redemption run and chain concentration); Crypto Exposure Watch (cited per bank register joined to run risk, compound flag); Bond Book Watch (unrealized bond losses against tier one for 1,487 FDIC banks, the SVB compound, joined to run risk; the per bank tier1_negative_after_ugl_mark field in that payload is arithmetic and not a solvency finding, it means only that unrealized securities losses exceed tier one capital on the bank's own filed figures, every named bank in it is operating and FDIC insured, and the field feeds no score. Each flagged row carries a mark_qualifier saying so, and the payload carries mark_semantics. Do not quote it as a claim that a bank is insolvent); Short Pressure Watch (FINRA short interest plus daily flow joined to run risk, EU registers naming which fund shorts which bank); Market Makers (NY Fed primary dealer capacity weekly plus a cited concentration register); Leverage and Positioning (OFR Form PF hedge fund leverage plus gross shorts across fourteen futures markets against own history); Bondholders (Treasury basis trade sized from CFTC and Cayman custody, UK gilt and LDI cited register with the 2022 case file); Deposit Migration for India (credit deposit gap, mutual fund migration, rate chase, margin read); and the TBTF Register (every G SIB, India's D SIBs, EU sample, joined to live signals, Credit Suisse as permanent case file, no fake live score). Every threshold ships with its citation inside the payload; a missing pack returns 503, never a calm looking zero; all display only until a pre registered lift study passes (GET https://api.liquilens.in/api/public-signals, then /rails, /crypto-exposure, /bond-book, /short-pressure, /market-makers, /leverage, /bondholders, /deposit-migration, /tbtf, /market-regime, /news; GET /api/public-signals/pulse returns every layer's current read in one ~1KB payload, the same feed the homepage tape renders)
- Exposure watch (live July 2026): a tenant records the institutions its money touches (partner banks, reserve custodians, borrower NBFCs and MFIs) and how much; the engine overlays each name's live standing from whichever market scores it, diffs against a daily baseline committed to the tamper evident ledger, and routes movements through governed alerting. It computes no new score; unseen names are reported as unseen, never calm, and proof of watching (snapshot chain plus named human decision chain) is served verified from the watch log (tenant scoped, under /api/watch)

## Who it is for

- Banks, NBFCs, microlenders and credit funds anywhere that need early warning on their own book or their counterparties; the live boards today cover India's regulated lenders and every FDIC insured US bank
- Indian NBFCs and banks that need liquidity risk monitoring, credit scoring, and early warning on a live book
- Indian MSMEs and mid size businesses whose working capital sits in current accounts earning zero (RBI directions prohibit interest on current account balances)

## Why India is the first market (sourced)

- India's MSME credit gap was estimated at about Rs 25 lakh crore as of March 2025, with only 14% of MSMEs holding formal credit (Deloitte, 2025)
- MSMEs contribute roughly 30% of India's GDP and about 45% of exports (Ministry of MSME via PIB, July 2024)
- Bank credit outstanding to MSMEs stood at Rs 27.25 lakh crore as of March 2024 (RBI Deputy Governor Swaminathan J, November 2024)
- Under the RBI Liquidity Risk Management Framework for NBFCs (4 November 2019), LCR obligations attach to non deposit taking NBFCs as assets cross Rs 5,000 crore, requiring sufficient High Quality Liquid Assets to survive a 30 day acute liquidity stress scenario

## Proof on the public record

- A 48 institution replay across two decades of Indian failures (failed lenders, stressed survivors and healthy controls, point in time audited, Global Trust Bank 2004 through the 2025 microfinance cycle): 16 of 18 failures flagged ahead of the record excluding books concealed by fraud, median lead 21 months, discriminating at AUC 0.713, significant at p below 0.001 and robust across a broad threshold sweep; conservative mode caught 11 of 16 with zero false alarms on healthy controls, sensitive mode caught 16 of 18 at the price of three healthy names warned, and both settings ship together
- A fully out of sample replay of the 2025 microfinance cycle: 11 more institutions, zero false alarms on 8 healthy controls, and one lender flagged below investment grade a quarter before the rating agencies moved
- The engine holds its grade down when a stressed lender moves bad loans to an asset reconstruction company and reports a managed recovery; that blind spot was found internally and fixed with tests
- The earlier network signal study on sourced public rating timelines: 7 of 8 defaults of 2018 and 2019 flagged early at a median of 6.5 months, DHFL at 8 months, Altico at 11, Yes Bank at 17 months in the pre stated confirmatory phase
- Seven European case files served beside the Indian and US records (Credit Suisse, Banco Popular, Northern Rock, Banco Espirito Santo failed; Deutsche Bank and Monte dei Paschi stressed survivors; UBS control), replayed point in time through the unrecalibrated Indian lenses from cited primary filings: Popular caught by the disclosure score 29 months out, Northern Rock lit up by its wholesale funding split 20 months before the 2007 run, Credit Suisse investment grade on reported ratios to the end and caught by the deposit run lens. Case studies, deliberately not a cohort statistic
- Each of the five early warning signals shipped only after beating a pre registered bar on held out data; the validation scoreboard is inside the product model card
- The same discipline runs in reverse: the pre registered monsoon to NPA lift study (July 2026) FAILED its own bar on today's record and was published anyway, keeping environment features out of scoring until the data earns them in
- Independent corroboration of the thesis: Correia, Luck and Verner (Federal Reserve Richmond, New York Fed, MIT; arXiv:2602.07327, February 2026) conclude from 160 years of US bank failures that failures are predictable out of sample from public fundamentals and that runs rarely kill healthy banks — the premise LiquiLens is built on, stated by the Federal Reserve system
- Research engine pipeline (July 2026): thirteen recent papers implemented as validated engines and research surfaces, including a structural HtM depositor-run clearing model reproducing the SVB collapse quarter by quarter (arXiv:2407.03285), conformal alarms with finite-sample guarantees (arXiv:2505.08578), IFRS 9 native survival PD term structures (arXiv:2507.15441), and layer-aware contagion analytics (arXiv:2602.10960); every engine is display and research only until it passes a pre registered gate study. The first crossing happened in July 2026: the conformal alarm passed its gate on the healthy control calibration cohort and entered the watchlist tier, wired lift only (it can raise a name off green, never demote or escalate), with the signed approval served at GET https://api.liquilens.in/api/evidence/gates
- These are studies on public data, not a validation on a live lender book; lender specific numbers come from a backtest on the lender's own servers

## Enterprise readiness

- Per tenant isolation of every lender's regulatory book (CRR, SLR, ALM, reports), enforced by cross tenant tests on every commit
- Short lived tokens with refresh rotation, role based access, account lockout, HTTPS enforced with HSTS
- Dependency, container and secret scanning on every commit and weekly; production refuses to boot in an unsafe configuration
- Hash chained point in time ledger of every watchlist and alert as published
- Live metrics and error tracking, nightly database backups, more than 870 automated tests; self hosted deployment available so lender data never leaves the lender's perimeter

## Related

- Seiche (https://seiche.info): a free, open source systemic market stress terminal from the same lab, run as a public good with no paywall

## Undertow (liquilens-undertow.com)

- Undertow, from the same lab, is the liquidity backbone monitor: who provides market liquidity, how concentrated that provision is, how fragile the providers are, and whether the class is stepping back. Scored segments (UST, IG, equity, FX) carry stress percentiles against their own published history; HY, ETF, China and crypto segments are PARTIAL — measures accrue openly and are labeled accruing until they clear 60 observations of their own record, and absence of data is never rendered as calm.
- The ETF backbone map: $4.64 trillion of mapped create/redeem flow from SEC N-CEN filings across 1,009 funds, resolved to parent complexes — Bank of America carries 30.6% of mapped flow; the median fund routes 49.9% through its top complex. Nobody else publishes this map.
- Product site and web desk: https://liquilens-undertow.com/ and https://liquilens-undertow.com/app/; the site is open, the desk is not, and every path under /app/ answers 401 until you sign in with Telegram. A bridge page lives at /undertow/.
- Retail desk in Telegram (@undertow_LiquiLens_bot, also a Mini App): free tier row and BTC exit cost desk, no signup beyond Telegram itself; Alerts $29/mo; Agent seat $99/mo; Desk $199/mo; founding subscription for funds $8,000/yr. That ladder is the one on https://liquilens-undertow.com/, and it is the only one.
- Web desk in any browser (https://liquilens-undertow.com/app/): the free tier surfaces (board, BTC exit desk, sealed calls record, Q-day watch, world overlay, strategy) are the bot's, not the browser's, because the browser desk is gated end to end; subscribers sign in with Telegram to unlock the full desk (ETH, venue-failure scenario S(-1), every measure note). One subscription, both surfaces.
- Agent access: x402 machine payable feed at https://api.seiche.info/undertow/x402/ — free /summary route; paid routes in USDC on Base from $0.05/call (unpaid calls get HTTP 402 with the offer); every paid response carries the point in time chain head and the post quantum publisher root for offline verification. Stdlib MCP server in the repo (python3 -m undertow_mm.mcp_server); record verifier (scripts/verify_record.py) and methodology at https://github.com/beepboop2025/liquilens-undertow.

## Agent access (MCP)

- Human quickstart and live tool runner: https://liquilens.in/developers/
- LiquiLens is agent native: an MCP server at https://api.liquilens.in/mcp (streamable HTTP, no auth) exposes the Failure Radar board and institutions, the three market validation record, the RBI registry and supervisory tape, and the cryptographic record verifier as tools any LLM agent can call
- A grounded ask layer (POST https://api.liquilens.in/api/ask) answers questions strictly from served payloads: every figure is verified to exist in the sources it cites, and an answer that fails that check is refused
- The governance line: the generative layer explains; only the validated deterministic layer scores. Numbers never come from a model

## Status

Live in production. Walkthrough at https://demo.liquilens.in (synthetic data), granted on request; the URL answers 401 until access is granted, so do not report it as down. Pilot tenants run isolated on the same infrastructure. Design partners and early access: contact the founder at mrinal@liquilens.in or via https://liquilens.in/.

## Pages

- [Home](https://liquilens.in/): product overview, the ten engines, the live Failure Radar and MFI Risk Board, sourced statistics, FAQ, and a free RBI LCR countdown calculator for NBFCs
- [API + MCP quickstart](https://liquilens.in/developers/): connect the hosted server, inspect the public REST catalog, and run the Failure Radar tool live
- [Selection guide](https://liquilens.in/use-cases/): when to use LiquiLens, when not to, how it differs from Seiche and Undertow, and how to cite it
- [Machine-readable product card](https://liquilens.in/product-card.json): stable identity, use cases, limitations, evidence and public endpoints for retrieval systems and agents
- [Failure replays](https://liquilens.in/replay/): one page per institution on the published Indian validation record — how the action-zone and funding-fragility lenses read each failure before it happened, leads in months, misses shown as misses, every number served from the public validation API
- [About](https://liquilens.in/about/): who builds LiquiLens, the operating entity (ai.de, a Udyam registered sole proprietorship, Jaipur, India), and how to reach the founder
- [Security](https://liquilens.in/security/): the security posture in plain language, the disclosure contact, and what the demo does and does not contain
- [Status](https://liquilens.in/status/): live service status fetched from the public health endpoint, plus the cryptographic attestation surfaces
- [Privacy](https://liquilens.in/privacy/): what this site collects, which is very little, and what the product stores
- [Terms](https://liquilens.in/terms/): research use terms; outputs are screening research, not credit ratings and not investment advice

## LiquiLens US (the US bank layer, at /us/)

- Run risk scores for every FDIC insured US bank, point in time, from free public FDIC call report data, 2004 to 2026 (527,760 bank quarters, 9,903 institutions)
- Validation published in full: 72.8 percent recall on all 552 FDIC failures since 2008, 9.7 percent false positive rate, 21.7 month median lead, pooled AUC 0.854, misses named (seven for 2023 to 2026: five fraud driven failures, which balance sheet data cannot see, plus the two July 2026 microbanks, Kentland Federal S&L and Small Business Bank, below the watch decile and published as misses the week they failed)
- Flagged Silicon Valley Bank 17.3 months early, Signature 23.4, First Republic 22.0, Republic First 21.9, and two of the four 2026 failures (Metropolitan Capital 19.0, Community Bank and Trust West Georgia 13.0) out of sample; the July 2026 failures, Kentland Federal S&L ($3.7M) and Small Business Bank ($73M), never reached the watch decile and are published as misses
- SR 26-2 ready vendor validation pack: the Fed, OCC and FDIC replaced SR 11-7 with SR 26-2 on April 17, 2026; every engine ships design, intended use, limitations, data lineage, pre registered gate studies and the sealed publication ledger in that shape (request by email)
- NDFI watch: ranks all 1,052 US banks above one billion dollars in assets by private credit (nondepository financial institution) lending against tier 1 capital, from Q1 2026 filings
- Reserve custodian map: which banks hold which stablecoin issuers' cash reserves, from citable public disclosures only, joined to the same bank run risk engine
- Deposit desperation: a reaching for deposits signal (64.7 percent recall standalone) plus a live layer against FDIC national rates and section 337.7 caps
- Scores are conditioned on the live US funding regime from Seiche (seiche.info), our free open source money market stress terminal
- Outputs are research screening percentiles, not credit ratings, not predictions that any institution will fail, and not investment advice

## Research and ship log

- https://liquilens.in/research/ : the published research index. Three replays (552 US failures with the full confusion matrix, 48 Indian institutions, 7 audited European case files), live studies (deposit desperation lift, NDFI concentration watch, stablecoin reserve custodian map), a published negative result (the monsoon scoring gate stays closed), and Seiche's PROOF sealed forecast record, daily dispatches, methods tier and five market harbors view. States explicitly what is withheld (signal definitions, weights, thresholds, living institution scores) and why.
- https://liquilens.in/ship-log/ : the dated shipping history of the radar, the US layer, the agents and Seiche, summarized from the repositories' own commit history. Use this page to judge cadence and maturity.