Four AI agents debate a distressed-credit situation — leverage, recovery waterfall, fulcrum security, tail risks — and write the IC vote memo. Point it at any deal in a YAML file. Every number comes from deterministic, unit-tested Python, not the model — so the math is auditable, not hallucinated. Model-agnostic (Claude / GPT / Grok / local Ollama). MIT licensed.
The open-source "AI investing" projects that exist — ai-hedge-fund (~61k★), TradingAgents (~92k★), FinRobot (~7.5k★) — are all equities/trading. None does distressed credit or restructuring: recovery waterfalls, fulcrum-security identification, covenant analysis. This is the first to bring the multi-agent pattern to the liability side of the balance sheet.
The reason it isn't just another prompt wrapper: the LLM never does the arithmetic. Leverage, the pari-passu recovery waterfall, attachment/detachment in turns of EBITDA, the fulcrum security, asset coverage, the maturity wall — all deterministic, all unit-tested. The model brings judgment; the code keeps the numbers honest. You can see all of it for free, with no API key:
quantai-credit validate my_deal.yaml # instant cap-structure snapshot, no LLM
Analysis and education tool, not investment advice. It makes no claim to beat any benchmark, and it doesn't replace a data terminal (Bloomberg, Octus/Reorg, 9fin) — it's a free, local, auditable analysis layer for the cap-structure work you'd otherwise build by hand.
git clone https://github.com/RahulModugula/quantai-dashboard.git
cd quantai-dashboard
python -m examples.distressed.demoPrints a full 4-agent credit committee memo on ATI Physical Therapy's April 2023 Transaction Support Agreement — an out-of-court loan-to-own restructuring — using only the Python standard library. It's the bundled worked example; the next section runs the committee on your own situation.
The ATI case is analyzed at the April 2023 entry point, not in hindsight. The position later closed as a $523.3M take-private in August 2025 (~11.2x LTM Adj EBITDA) — shown as directional context for the thesis, not as a forecast the tool produced. See docs/ARCHITECTURE.md for design notes.
The committee isn't hardcoded to ATI — point it at any distressed situation in a YAML file and it writes the IC memo:
pip install -e . # lightweight: litellm + pyyaml + rich, no ML stack
quantai-credit new my_deal.yaml # scaffold an annotated template
# ...fill in the cap structure, timeline, metrics, and risks...
quantai-credit validate my_deal.yaml # free pre-flight: computed snapshot + sanity checks
quantai-credit run my_deal.yaml # 4-agent committee → my_deal_memo.md + my_deal.json
quantai-credit list # show bundled example situationsA situation file is just the cap stack, a timeline, operating metrics, and the
risks you already see — no code. validate computes leverage, coverage,
attachment/detachment, and the maturity wall instantly and free; run adds the
4-agent debate and vote, writing both a human-readable memo (_memo.md) and a
machine-readable result (.json). Start from
TEMPLATE.yaml or copy a bundled example:
| Situation | Structure it teaches |
|---|---|
ati_2023.yaml |
Loan-to-own via a 2L PIK convertible fulcrum (out-of-court TSA) |
serta_2020.yaml |
Non-pro-rata uptier / liability management — inside vs. outside the majority |
hertz_2020.yaml |
Asset-coverage with a bankruptcy-remote fleet-ABS silo (Chapter 11) |
Each is sourced from public filings, with approximate figures marked inline. Adding another is pure YAML — a great first contribution.
No LLM key?
python -m examples.distressed.demoshows a complete sample memo with zero setup. To run live, set any LiteLLM-supported key (ANTHROPIC_API_KEY,OPENAI_API_KEY,OPENROUTER_API_KEY) or pointQUANTAI_AGENT_MODEL=ollama/llama3at a local model for zero cost.
An AI distressed-credit committee (examples/distressed/). Four agents debate a
restructuring and write an IC-style vote memo. They call deterministic Python tools
— the math doesn't vary by temperature — and hand structured briefs to each other:
- CapStructureAgent — leverage, coverage, fulcrum security, recovery waterfall (base/bear/bull)
- SituationAgent — docket/timeline events, catalysts, structural vs. noise, information gaps
- CreditRiskAgent — devil's advocate: stresses every assumption, enumerates tail and process risks
- CreditCommitteeAgent — the vote memo: instrument, sizing, target, catalyst, downside, conditions
You describe a situation in YAML — cap stack, timeline, operating metrics, known
risks — and run quantai-credit run. No code.
Every quantitative claim traces to an audited function in
credit_tools.py, not to the model:
- One canonical claim per tranche. Face value is used consistently across the waterfall, breakeven, and attachment/detachment, so the tools never disagree about the same tranche. PIK accrual is opt-in and applied identically everywhere.
- Preferred/equity is not debt. Debt/EBITDA is computed over funded debt only; preferred sits in the recovery waterfall but never inflates leverage.
- Fulcrum is correct at the boundary. The fulcrum is the most-senior impaired claim — including when enterprise value lands exactly on a tranche boundary.
- It refuses to fake the hard cases. The generic waterfall is a pari-passu,
going-concern (EV = EBITDA × multiple) model. When a situation is a non-pro-rata
uptier (Serta) or an asset-backed / bankruptcy-remote silo (Hertz),
validatesays so loudly rather than printing a confident wrong number — and ships anasset_coverage_ratio/collateral_recovery_pctprimitive for the silo case. Honesty about model limits is a feature.
This design directly answers the best-evidenced critique of LLMs on money — they confidently get financial arithmetic wrong (FinanceBench found GPT-4-Turbo wrong or refusing on ~81% of open-book financial questions). Here the model never touches the arithmetic.
Not a textbook example — a real situation analyzed at the decision point:
| Entry | April 11, 2023 — Transaction Support Agreement; 2L PIK convertible, loan-to-own |
| Cap structure (pro forma) | $575M funded debt / 85.8x on $6.7M FY2022 EBITDA; $165M preferred excluded |
| Thesis | PT wage normalization → EBITDA recovery → fulcrum equity conversion |
| System vote | BUY — APPROVE WITH CONDITIONS, 1.0–1.5% AUM |
| Outcome (context) | Aug 1, 2025: $523.3M TEV take-private at ~11.2x LTM Adj EBITDA |
Capital structure, operating metrics, and timeline are sourced from public filings (ATI 10-K FY2022, 10-Q Q1 2023, 8-K 04/21/2023). The August 2025 outcome is shown as directional context, not as a prediction the tool made.
The committee runs on three dependencies — litellm, pyyaml, rich. No ML
stack, no torch, no dashboard:
pip install -e . # adds the `quantai-credit` command on your PATH
quantai-credit list
quantai-credit validate examples/distressed/situations/ati_2023.yaml # free, no keyCI proves this: a dedicated job installs the credit committee without the ML
stack and runs validate on every push.
The same BaseAgent loop also drives an equity-research pipeline — a second proof
that the architecture is asset-class-agnostic. It's a fuller "batteries-included"
trading playground and is not required for the credit committee. It lives
behind the equity extra and pulls the ML/data stack (torch, scikit-learn,
xgboost, Dash, FastAPI, ...):
pip install -e ".[equity]" # or: docker compose up --build
make seed && make train && make run # http://localhost:8000What it includes: QuantAgent → NewsAgent → RiskAgent → PortfolioManager over a
walk-forward ML ensemble (RF + XGB + LightGBM + LSTM, no lookahead bias), SHAP
explainability, backtesting with Monte Carlo CIs, a FastAPI service, and a Plotly
Dash dashboard. See docs/ARCHITECTURE.md for the full tour.
make test-credit # credit committee only (no ML stack)
make test # everything, with coverageThe credit tools are covered by dedicated suites —
test_distressed_credit.py,
test_credit_correctness.py (regression tests
pinning the fulcrum-boundary, tool-consistency, debt-vs-preferred, and
asset-coverage fixes), plus snapshot, situation-loader, and second-case (Envision)
suites. They verify the waterfall, fulcrum, leverage/coverage, attachment/detachment,
and the bundled situations against sourced numbers. The broader repo (~380 tests)
also covers the equity reference build.
examples/distressed/ # THE CREDIT COMMITTEE (the core)
├── models.py # Situation, CapitalStructureTranche (debt vs preferred)
├── credit_tools.py # leverage, coverage, waterfall, fulcrum, asset coverage
├── snapshot.py # free `validate` snapshot + structural warnings
├── agents.py # CapStructure, Situation, CreditRisk, CreditCommittee
├── run.py # the `quantai-credit` CLI
├── situations/ # ati_2023, serta_2020, hertz_2020, TEMPLATE (YAML)
└── demo.py # zero-dependency terminal demo
src/ # EQUITY REFERENCE BUILD (optional `[equity]` extra)
├── agents/ # shared BaseAgent + equity multi-agent layer
├── models/ data/ backtest/ trading/ advisor/ api/ dashboard/
- Public data only. Recovery and priority ultimately turn on credit agreements, indentures, and intercreditor terms that aren't always public. Treat this as pre-diligence triage and a structured-thinking framework — feed it your own documents when you have them; it structures the analysis, it doesn't source private data.
- Multi-agent debate is not magic. For genuine adversarial value, run the agents on different models via LiteLLM rather than four clones — homogeneous "debate" can manufacture false consensus.
- LLMs can still hallucinate the narrative even when the math is fixed. The deterministic tools constrain the numbers, not the prose.
- Not investment advice, no performance claim, no reliance. The equity reference build likely does not beat buy-and-hold after costs — expected, and consistent with the EMH for liquid US equities.
MIT