# eqlx.org

> Open-source deterministic tools for AI agents in credit and finance, from eqlx (https://eqlx.ai). The agent decides what to check and when; the tool computes the number. Every tool is deterministic (no language model inside the calculation), callable by agents, verified against published tables or official samples, runs locally, and is Apache-2.0.

## Projects

- [eqlx-credit](https://eqlx.org/credit/): Python library for PD (probability of default) model validation: AUC/Gini with DeLong confidence intervals, the ECB test of AUC decline since initial validation, Jeffreys and binomial back-tests (optionally with correlated defaults), traffic lights, Hosmer-Lemeshow, normal test, Spiegelhalter, cohort default rates and long-run average (EBA/GL/2017/16), PSI, HHI, calibration (central tendency, Tasche CT/AR curve), master scale, Pluto-Tasche and van der Burgt for low-default portfolios, and Basel IRB risk weights reproducing Basel II Annex 5. Includes an Agent Skill "annual-pd-validation". Install: `pip install eqlx-credit`. Source: https://github.com/eqlx-ai/eqlx-credit
- [eqlx-ebelge](https://eqlx.org/ebelge/): Python library and read-only MCP server for Turkish e-Fatura / e-Arşiv invoices (UBL-TR) and e-Defter journals and ledgers (XBRL GL): clean data, trial balance (mizan), and checks for VKN/TCKN check digits, invoice arithmetic including tevkifat and stopaj, duplicate invoices, and unbalanced journal entries. Checked against GİB's public sample packages. Install: `pip install "eqlx-ebelge[mcp]"`. Source: https://github.com/eqlx-ai/eqlx-ebelge
- [eqlx-aml](https://eqlx.org/aml/): Python library and read-only MCP server for deterministic money-laundering pattern detection in transfer data (CSV, generic or IBM AML layout): time-respecting loops, fan-out, fan-in, gather-scatter, scatter-gather, pass-through and structuring under a given threshold, plus account profiles and time-respecting fund tracing. Every alert lists accounts, time span and input rows. Each detector matches a brute-force enumeration of its rule; on IBM's public AML data (HI-Small, 5,078,345 transfers) it finds 140 of the 144 planted attempts that fit a detector's rule. Install: `pip install "eqlx-aml[mcp]"`. Source: https://github.com/eqlx-ai/eqlx-aml
- [eqlx-ifrs9](https://eqlx.org/ifrs9/): Python library and read-only MCP server for deterministic IFRS 9 / TFRS 9 expected credit loss on a loan file: monthly PD term structures (from a 12-month PD or a rating transition matrix), one-factor scenario shifts with reversion to the long-run PD, repayment schedules and undrawn limits, 12-month and lifetime ECL discounted at the effective interest rate, stage allocation with every rule that fired (relative and absolute PD tests, threefold backstop, 30 and 90 days past due, forborne, watch list, low credit risk exemption, probation), probability-weighted scenarios with per-scenario sensitivity, the provision matrix for trade receivables, and a Turkey (BDDK provisioning regulation) profile with five groups, restructuring and borrower-level contagion. Since 0.2 also the work around the number: reconciliation of a lender's reported stage and allowance with the recalculation (5% / 10% thresholds of the ECB asset quality review, gap split into staging and measurement), a scenario response curve, a challenger with the rating migration observed between two loan files, the movement of the loss allowance and gross carrying amount between two reporting dates by stage (IFRS 7 35H-35I) with the change split into drivers, staging quality indicators, a backtest of defaults that separates forecast error from model error, and the individual assessment of a credit-impaired debtor (gone concern, steady-state and two-step going concern, market price, sharing by seniority, probability-weighted scenarios). It does not estimate PD/LGD or forecast the economy; those are inputs. Reproduces IFRS 9 Illustrative Examples 8, 9, 10 and 12, IFRS 7 IG20D, the ITG December 2015 scenario example (CU92) and the worked examples of the ECB Asset Quality Review manual (section 4.6) and the ECB guidance on non-performing loans (section 7.4.3). Install: `pip install "eqlx-ifrs9[mcp]"`. Source: https://github.com/eqlx-ai/eqlx-ifrs9
- [eqlx-bankmath](https://eqlx.org/bankmath/): Python library and MCP server (18 read-only calculators, no files read or written, no dependencies) for bank product arithmetic in Türkiye in exact decimal arithmetic: instalment loan plans with the levies on interest (KKDF, BSMV) looked up by product and disbursement date with their legal basis, the annual cost rate as the consumer credit regulation defines it (360-day year, twelve equal months), early repayment, partial prepayment, late interest, largest loan for an instalment; daily-balance interest for retail overdraft (KMH), revolving lines, current accounts and spot loans; credit card minimum payment, rate ceilings and statement interest; bill discounting; variable-rate plans; letter of guarantee commission under explicit conventions; checks of fees against regulated ceilings. Instalment plans reproduce the worked examples in the annexes of the Turkish consumer credit and housing finance regulations to the kuruş; products with no official worked example (cards, overdraft, guarantees) are checked against their definitions, and every result lists the conventions it used. A calculation, not an offer or legal advice; rate tables are dated. Install: `pip install "eqlx-bankmath[mcp]"`. Source: https://github.com/eqlx-ai/eqlx-bankmath
- [eqlx-portfolio](https://eqlx.org/portfolio/): Python library and read-only MCP server (11 tools) for deterministic credit portfolio risk on a loan file (CSV: exposure, default probability, loss given default, optional borrower and sector) with settings (JSON): the loss distribution of the whole book, expected loss, value-at-risk with a distribution-free band, expected shortfall, economic capital, each borrower's and sector's share of the expected shortfall, concentration, and a what-if for new loans valued on the same scenarios. Two methods: a seeded Monte Carlo simulation of the asset value model (one common factor, the Basel corporate correlation, or sector factors) and CreditRisk+ in closed form with a numerically stable expansion. Since 0.2 it also prices a new loan: cost components (funding, expected loss, operating cost, capital), the hurdle rate at which the loan earns the target return, RAROC and yearly value added, with the capital it ties up measured from the portfolio itself, from the formula behind the Basel risk weights and from a fixed risk weight; pricing follows the scheme of Engelmann and Pham (2020) and is verified by its closed form and hand arithmetic, not by published tables. It estimates nothing: default probabilities, loss given default, correlations, funding costs and target returns are inputs; the hurdle rate is a break-even point, not a recommended price. Portfolio risk is checked against the worked examples of the CreditMetrics Technical Document (1997) and the 25-obligor example of the CreditRisk+ document (1997), closed forms and brute-force simulation; the same seed gives the same numbers across NumPy versions. Portfolio runs are one-period and default-mode. Source: https://github.com/eqlx-ai/eqlx-portfolio

## Evidence

- [How the tools are tested](https://eqlx.org/testing/): every tool's headline checks against published examples (Basel, IASB, ECB, Turkish regulations, GİB, CreditMetrics, CreditRisk+, papers), closed forms, brute force and simulation, and the errors found in official documents while reproducing them.

## In preparation

- TR-CreditBench: a public Turkish test set measuring which credit-risk tasks a language model can do on its own and which it must hand to a deterministic tool.

## About eqlx

- eqlx.org is the open-source site of eqlx (https://eqlx.ai), an İstanbul company that builds AI that thinks and works like your expert: machine-learning models, small language models trained for one domain, and agentic AI systems, for financial institutions, industrial and service companies. eqlx's own site and its llms.txt: https://eqlx.ai/llms.txt
- Open models: https://huggingface.co/eqlx-ai · Code: https://github.com/eqlx-ai

## Contact

- research@eqlx.ai