A public finance research hub for capital markets, funding, banking, market structure, company and industry-chain research, and reproducible financial analysis. Deep Funding, the existing agent-to-agent funding-matchmaking sandbox, is the first project, not the boundary of the repository. New work is researched directly in public.
P Bank and P Street are internal research initiatives; PDD is an internal due-diligence source. Only selected material explicitly cleared for public release may be cited or adapted. Do not publish private URLs, reports, datasets, positions, counterparties or contact details. Distinguish facts, inference and opinion; include dates and data rights. No individualized investment advice, deals, outreach, or live transactions.
Humans and AI agents: start with AGENTS.md or the call for contributions (#9). Finance-related repos, papers, projects, experience reports and reproductions are welcome. See the community scope.
- paper · Paper Discussions: Papers on finance, markets, funding or industrial organization, including Academic Papers and its recommendation thread.
- repo · Repo Discussions: Open financial software and data tooling with license and provenance.
- report · Report Discussions: Original company, industry-chain and due-diligence reports.
- website · Website Discussions: Public filings, regulators, official data and finance education.
- project · Project Discussions: Runnable finance research and the existing Deep Funding sandbox.
Browse the expandable finance repo directory. Bank, PE, VC, securities, exchanges, funds, insurance, and financial data. Shared repositories appear under each relevant use without duplicating their source pins.
Original bank and markets routes: Corporate Banking · Investment Banking · FICC · Asset Custody. Cross-listed projects have a different use note for each topic. See the dated catalog update.
Earlier: Banking / IB / PE / VC repository map and dated investigation report. Four named bank organizations:143 public non-forks;45 selected cross-ecosystem guides, with pinned sources, license/state boundaries and separate reference-only candidates. These are source reviews, not143 banking applications or executed financial models.
Research reading routes: Corporate Finance repositories, FICC / curves, pricing and risk, and how to study and reproduce a repo. Public-source guides, not production endorsements; the existing Deep Funding application is unchanged.
Companies want the right capital. Capital wants the right company. Let their agents meet first.
Open-source, evidence-aware Agent-to-Agent funding matchmaking. Two top-level agents, eight scoped sub-agents, one inspectable match protocol. A runnable fictional-data sandbox, local private research database and sourced public knowledge library. Not an investment recommendation or autonomous dealmaker.
ModelScope deployment candidate: hybrid public Studio package and current deployment status / historical-document corrections. ModelScope login and a verified Studio are still required; this is not yet a live Studio.
3D product-concept film: Agent City source, preview and MP4 export.
Watch the 28-second 1080p film online.
The web master is versioned under apps/web/public/film/; the original master remains local.
28 seconds, two transparent hubs, bidirectional Agent flow and a synthetic A2A shortlist.
This illustrates the vision; it is not evidence of real outreach or funding outcomes.
- SQLite / SQLAlchemy / FastAPI: twelve models, company and funding-provider forms, funding preferences/needs, saved matches, four-stage Agent run history, audits and lightweight subscription records.
- Funding RAG Graph: 18 sourced institution/program seeds, classification nodes and templates. Verify current eligibility separately; unknown tickets and terms stay unknown.
- Compliance RAG Graph: sourced references and evidence-request templates, with allegations, responses and findings kept separate. No blacklist or guilt by association.
- Data Explorer: tables, detail JSON, counts, search, sorting, filters and persistent forms.
- Knowledge Explorer: Funding and Compliance.
Cloudflare serves public knowledge and the existing fictional D1 demo. The new private SQLite database runs locally, not on Cloudflare. The public Database page links to that local workspace. Hosted private authentication and cross-device database sync are not implemented.
Python 3.11+, Node 24. No model key required. Windows PowerShell:
py -3.11 -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r backend/requirements.txt
npm.cmd ci
.\.venv\Scripts\python.exe scripts/build_knowledge.py --check
npm.cmd run build
.\.venv\Scripts\python.exe scripts/seed_database.py
.\.venv\Scripts\python.exe scripts/serve_database.pymacOS/Linux: create and activate a virtual environment with python3, then use python instead of the Windows executable and npm instead of npm.cmd.
Open http://127.0.0.1:8793/data-explorer. Swagger: /docs; ReDoc: /redoc. Default file: data/deep_funding.db, ignored by Git. The seed is idempotent and synthetic: three founders/companies, ten fictional providers and persisted matches. No real investor mandate is inferred from a public source record.
Use the Users / Founders / Companies / Funding providers / Funding needs tables to add profiles. Generate match persists the result, four Agent runs and an audit. The engine is deterministic rules-v0.3.0 with source-linked retrieval, not an LLM or a learned ranker. Scoring reports known-dimension coverage, not funding probability.
The service is single-owner loopback-only; do not expose it through a tunnel. Close the service before making a protected backup of the database. Do not commit the database or private exports. Existing Worker/D1 and RSI stores remain separate.
.\.venv\Scripts\python.exe -m pytest backend/tests -q
.\.venv\Scripts\python.exe scripts/benchmark_database.py
$env:PYTHON_BIN=(Resolve-Path .venv/Scripts/python.exe).Path
npm.cmd run test:database-uiDetails: architecture / security, V0.3 acceptance, changelog.
Your data. Your benchmark. RSI workspace combines Founder RSI, Funding RSI and an explainable, deterministic benchmark engine. GUI, keyboard-first TUI and CLI reuse the same TypeScript application services. Private RSI files stay in browser memory or a local terminal store; no remote inference or upload is implemented.
- Founder RSI: mandate gates + scoped interaction history + portfolio similarity; explain every component and unknown.
- Funding RSI: eight cohort modes, configurable feature weights/bounds, missing-data coverage, descriptive thresholds, empirical percentiles and append-only versioned observations.
- Shenzhen-first data foundation: strict JSONL/schema and source workflow, zero verified real Shenzhen profiles at this release. Three Shenzhen-labelled examples are fictional.
- Terminal:
npm run cli -- --help, ornpm linkthendeepfunding --help(deepfunding.cmdwhen PowerShell blocks npm's script shim).
Guides: Founder RSI, Funding RSI, Benchmark, CLI, TUI, Private data.
Funding is not only money. Funding is the set of capital and resources that helps a founder or company reach its next stage.
Funding Explorer now covers nine groups, sixteen multi-label categories, eleven company-stage labels and ten versioned screening policies. Browse, search, inspect source provenance, download canonical JSON and run a full A2A preview. Import a provider into your private workspace to reuse the existing audited workflow.
Institution categories, capital instruments and non-financial resources are separate. Company stage and financing history are separate. Resource-only requests do not need a cash ticket. Bank/PE/growth policies require financial disclosures; early-stage programs do not universally require revenue or a launched product. Readiness is evidence coverage, not funding probability.
The starter catalogue contains 14 fictional fixtures and one partial YC scaffold, not 15 verified funding opportunities. YC terms, deadlines and incomplete mandate fields remain unknown. No application or introduction is sent. See the knowledge base, integration notes and V0.2 release.
- Open Deep Funding · Protocol and boundaries
- Click Run the A2A demo to create a private browser workspace, load one fictional company and five mandates, run both sides and inspect the leading match.
- Review the weighted decomposition, hard failures, evidence and A2A trace. Record an introduction request after human review. No message is sent.
- Company and Capital workspaces support creating/editing profiles, source evidence, audits, matching, JSON export and complete workspace deletion.
Node.js 24 and npm are required. No LLM key or paid API is needed.
npm ci
npm startOpen http://127.0.0.1:8791. Local D1 migrations and the frontend build run before the Worker. The startup fails if the port is occupied; choose another port with npx wrangler dev --port 8792 after the initial build.
npm run check
npx playwright install chromium
npm run test:e2e
npm run benchmarkFor frontend hot reload: keep the Worker running and use npm run dev:web on port 5191. Mutations must stay same-origin through the Vite proxy.
Company Agent Funding Agent
Information Information
Analysis Analysis
Audit Audit
Match -------- A2A Match Layer ------ Match
|
Explainable result + trace
|
Human introduction queue
(recorded, not sent)
apps/web: React, TypeScript, responsive routes and evidence/trace views.apps/api: Cloudflare Worker HTTP API, same-origin sessions and D1 persistence.packages/agents: eight deterministic agent responsibilities.packages/schemas: strict Zod inputs and typed output contracts.worker-configuration.d.ts: generated locally bynpm run types/npm run check, not committed.packages/a2a: versioned application-level messages. No external A2A-standard interoperability claim.packages/matching: hard constraints, evidence gates and configurable weighted fit.packages/knowledge: canonical taxonomy, validated profiles, readiness, compatibility adapters and policy screening.funding/: category-governed Markdown knowledge base and canonical JSON profiles.schemas/funding_profile.schema.json: generated structural JSON Schema; runtime validation also applies cross-field rules.packages/connectors: provider-agnostic optional explanation and external-data boundaries. No live providers configured.data/demo.ts: entirely fictional demo profiles, generated with an explicit date.
Information → Analysis → Audit → bilateral Match → decision → human request, with real CRUD, database writes and inspectable protocol events. Matching uses rules-2.0.0, not a hidden LLM. The optional explanation adapter accepts only bounded text, labels it INFERRED, and cannot change the authoritative verdict. It is tested using a stub and is not invoked by the hosted app.
Mandates can be edited through the UI. Hard filters always veto scoring. A match requires current shared evidence and a score of at least 75 for human review. The sample leading match is 96.67, calculated rather than forced to the brief's illustrative score. See matching policy.
Use fictional data only. The public demo has anonymous, cookie-bound workspaces, not verified company identities. A 256-bit HTTP-only session token is hashed before storage. Records are owner-scoped. Workspaces expire after seven days and are deleted by daily cleanup; users can delete earlier. Losing the cookie loses access. No account recovery, cross-device identity, partner invitations or real capital network exists in V0.2.
Core profile sharing requires explicit consent. Private notes and PRIVATE / NDA_REQUIRED evidence never enter A2A matches. PUBLIC means share-eligible inside this sandbox, not globally published. Existing snapshots remain in private history after consent withdrawal until deletion. All claims are PROVIDED or UNKNOWN; the app cannot self-assign VERIFIED. Source freshness is a configurable 180-day demo policy.
No uploads, due-diligence certification, investor verification, automatic emails, negotiation, investment transactions or guarantees. Notes and exports can contain the data you enter; do not commit them. Cloudflare's edge logs may include request metadata, but application logs never include profile bodies or session tokens. Database backups follow the hosting provider's retention; this is not a regulated data room.
See security model, API, acceptance scope and benchmark card.
npx wrangler login
npx wrangler d1 create deep-funding
# Set your own database_id in wrangler.jsonc; choose unique rate-limit namespace IDs.
npm run types
npm run db:remote
npm run deployUse a fresh database for each independent deployment. The checked-in deployment ID is a public resource identifier, not a credential. Keep credentials out of source control. GitHub CI checks types, unit tests, build and browser workflows. Deployments remain explicit CLI actions; CI does not have account secrets.
TODO.md separates V0.1 evidence from roadmap work: verified identities and consent, real provider adapters, semantic retrieval, signed inter-party messages, data rooms and financing-specific diligence. Pull requests should have clear scope, reproducible tests and an honest benchmark.
Code: MIT. See CONTRIBUTING.md, CODE_OF_CONDUCT.md and asset provenance.

