Fund Workbench

Open-source, evidence-first workflow tools for investors, accelerators, corp-dev teams, and operators.

Fund Workbench turns pitch decks, call notes, research, and recurring updates into traceable artifacts. It is designed around a simple rule: a model may propose structure, but it may not silently turn a claim into a fact.

The repository contains four connected tools:

Tool	Input	Output
Evidence Ledger	PDF, DOCX, Markdown, text, or pasted notes	Claims with source, status, quote, and confidence
Screening Brief	Evidence ledger + explicit thesis	Reviewable memo with transparent thesis routing
Diligence Planner	Evidence ledger	Prioritized questions tied to missing or conflicting evidence
Portfolio Update Comparator	Two lightweight company updates	Metric deltas, changed risks, asks, and follow-up queue

The web interface works immediately with a synthetic company and no API key. An optional OpenAI integration uses Structured Outputs to improve semantic extraction; deterministic code still validates source IDs, quotes, and evidence status after the model responds.

Important

Fund Workbench supports human review. It does not make investment decisions or replace legal, financial, security, commercial, or technical diligence.

Try it in two minutes

You need Python 3.10+ and uv.

git clone https://github.com/Hritikd/fund-workbench.git
cd fund-workbench
uv sync --extra dev
uv run streamlit run app.py

Open the local URL, click Load synthetic demo, then Build evidence ledger. The demo uses a fictional industrial-software company.

Generate all three portable demo artifacts from the command line:

uv run fund-workbench demo --output demo-output

No API key is required for either path.

Use AI-assisted extraction

Copy .env.example to .env, set OPENAI_API_KEY, then export it in your shell or load it with your preferred secret manager.

export OPENAI_API_KEY="your-key"
export OPENAI_MODEL="gpt-6-astra"
uv run streamlit run app.py

The optional integration uses the OpenAI Responses API with a Pydantic schema, following the official Structured Outputs guide. The request sets store=False. Your organization’s API configuration and data policies still apply; do not send confidential material without authorization.

Command-line workflows

Build a ledger from local documents:

uv run fund-workbench evidence \
  --company "Example Co" \
  --source-type company \
  notes/example-company.md notes/reference-call.md

Add --ai when a valid API key is present. Supported files are .txt, .md, .pdf, and .docx.

Create a memo from a ledger and thesis:

uv run fund-workbench memo evidence-ledger.json examples/thesis.json

Compare normalized portfolio updates:

uv run fund-workbench compare-updates previous.json current.json

Evidence statuses

Status	Meaning
verified	Supported by the supplied independent source; still subject to source quality
company_reported	Stated by company-controlled material
reported	Recorded in notes or an internal source
inferred	An explicit interpretation rather than a quoted fact
unknown	Required information is absent or explicitly unknown
contradicted	Sources disagree and require human resolution

Company material cannot mark its own claims verified. If an AI-generated quote is not found verbatim in the referenced source, Fund Workbench removes the quote and downgrades the claim to inferred.

Architecture

flowchart LR
    A[Decks / notes / research] --> B[Document extraction]
    B --> C{Extraction mode}
    C -->|No key| D[Deterministic rules]
    C -->|Optional AI| E[Structured Outputs]
    D --> F[Source-integrity validator]
    E --> F
    F --> G[Evidence ledger]
    G --> H[Screening brief]
    G --> I[Diligence questions]
    J[Recurring updates] --> K[Metric and risk comparator]
    K --> L[Follow-up queue]

Loading

The application does not include a database or telemetry. Uploaded files are processed in memory by the running process. See architecture, security and privacy, and design decisions.

Why this is not another deck summarizer

Summaries compress information but often erase provenance. Fund Workbench treats provenance as part of the output:

• Every material claim retains a source ID.
• Quotes must exist verbatim in that source.
• Company claims remain company-reported.
• Unknowns generate questions instead of synthetic answers.
• Thesis routing exposes matched terms and penalties.
• Update comparison converts changes into an operational follow-up queue.
• All artifacts export as JSON or Markdown and can enter another workflow.

Measure efficiency honestly

The project does not claim “10× faster.” Use the evaluation protocol to measure:

• time from source receipt to a reviewable brief;
• material claims captured;
• unsupported claims incorrectly promoted;
• contradictions surfaced;
• human editing time;
• questions accepted by the reviewer.

The sample regression tests check source integrity and key workflow behavior. They do not establish investment quality.

Develop

uv sync --extra dev
uv run pytest
uv run ruff check .

Or use Docker:

docker build -t fund-workbench .
docker run --rm -p 8501:8501 fund-workbench

Roadmap

• Editable claim review before export
• Configurable thesis templates
• Redaction helpers for sensitive documents
• CRM-neutral webhook output
• Reviewer calibration sets
• Additional model providers behind the same validated schema

Issues and pull requests are welcome. Please read CONTRIBUTING.md before contributing.

License

MIT © Hritik Datta