Turn filings, calls, peer metrics, and market reactions into a citation-bound investment thesis that can also say: no convincing thesis yet.
Try the live demo | Give Feedback | Quickstart | Methodology | Roadmap | Contribute
Watch the 91-second narrated AAPL walkthrough | Source event -> causal bridge -> peer evidence -> reverse DCF -> Bull/Bear/Judge
Most AI research tools are good at producing text. IC Copilot is designed to produce an auditable investment decision trail:
- Exact evidence: filing section, period, metric, unit, source tier, URL, and parser status.
- Causal reasoning: source event -> business driver -> KPI -> earnings/FCF -> valuation -> catalyst.
- Adversarial review: Bull, Bear, and Judge views with unresolved contradictions kept visible.
- Market-implied expectations: reverse DCF and implied operating assumptions when historical consensus is unavailable.
- A real ability to say no: boilerplate, stale facts, unmapped drivers, and unsupported LLM claims do not become high-conviction ideas.
The app is built for equity analysts, credit analysts, portfolio managers, and AI builders who want evidence before eloquence.
- Watch the narrated AAPL walkthrough, or open the hosted app. The AAPL Deep Initiation demo is loaded automatically.
- Read the one-screen IC Story: verdict, what changed, causal bridge, counter-thesis, valuation, and next action.
- Open an evidence drawer to inspect the exact citation and formula trace.
- Switch to NVDA, ORCL, BABA, TSLA, or GS to see a different research playbook.
The frozen demos require no API keys, make no paid provider calls, and contain no licensed full-text payloads.
flowchart LR
A[Entity] --> B[Sources]
B --> C[Claims]
C --> D[Business drivers]
D --> E[Peer read-through]
E --> F[Company model]
F --> G[Valuation]
G --> H[Price attribution]
H --> I[IC thesis]
I --> J[Monitor and post-mortem]
Every connection is scored separately. A weak link identifies the exact missing evidence instead of turning into generic uncertainty.
| Demo | Research lesson |
|---|---|
| AAPL | Source-backed cash-flow and margin thesis with model-derived scenarios |
| NVDA | Neutral-first capex and goodwill investigation with semiconductor peer checks |
| ORCL | Cloud/RPO growth versus data-center capex, financing, dilution, and cash conversion |
| BABA | ADR/FPI normalization, China segments, bilingual external context, and price-only capture |
| TSLA | Driver-specific peer metric read-through across automakers |
| GS | Bank/broker economics, equity lens, and credit lens |
| SPCX | Entity-resolution guardrail that stops unreliable analysis early |
git clone https://github.com/mwsp-code/IC-Copilot.git
cd IC-Copilot
python -m pip install -r requirements.txt
streamlit run app.pyWindows users can run the preflighted launcher instead:
.\scripts\run_streamlit.ps1python server.py --port 8501The lightweight server automatically tries ports through 8510 when 8501 is occupied.
python -m pip install -e ".[dev]"
pytest -qdocker build -t ic-copilot .
docker run --rm -p 8501:8501 ic-copilot- Fast Screening: 4 quarters, 2 annual reports, 4 calls, and the highest-ranked anomaly.
- Adaptive IC Research: default; 12 quarters, 4 annual reports, 12 calls, and five material changes.
- Deep Initiation: 20 quarters, 5 annual reports, 20 calls, and deeper contradiction work.
- Investigate This Event: a filing-, call-, metric-, or news-scoped workflow.
Ambiguous changes are neutral first. Constructive and adverse explanations are investigated side by side before direction is assigned.
SEC and issuer evidence remain the backbone. Official macro, market data, paid providers, Wisburg, and news can deepen context, but source tiers and licensing policies remain visible.
LLMs may plan registered sources, triage documents, draft structured extractions, compare evidence, and write an IC narrative. They cannot invent citations, targets, calibrated probabilities, or promotion decisions. Deterministic validation remains authoritative.
API keys are optional for the demos. For live work, save credentials through the OS keychain, environment variables, or Streamlit secrets. Never commit keys or licensed payloads. See Provider Configuration.
The repository ships a no-network benchmark over 25 research-integrity cases across AAPL, NVDA, BABA, TSLA, and GS. It checks event grounding, thesis grounding, counter-thesis handling, monitor/payoff readiness, and promotion integrity.
python scripts/run_benchmark.pyThis is an integrity benchmark, not a claim of investment performance. See Benchmark Methodology.
IC Copilot separates deterministic evidence collection from hypothesis generation, valuation, monitoring, and optional LLM synthesis. Provider failures are isolated, normalized, and source-labelled; missing values remain Unknown rather than becoming zero.
See Architecture and Research Methodology.
Useful first contributions include:
- Add a sector KPI playbook.
- Add an ADR/FPI profile.
- Add a canonical metric alias.
- Add a registered source adapter.
- Add a sanitized demo or benchmark case.
The UI and docs contain explicit contributor paths. Read CONTRIBUTING.md, browse issues labelled good first issue, or start a design question in GitHub Discussions after it is enabled.
IC Copilot is research software under active development. It does not execute trades and does not provide financial advice. Frozen demos illustrate methodology; refresh live sources before using any observation in an investment process.
See Roadmap, Security Policy, and Changelog.
If IC Copilot improves how you think about evidence, consider starring the repository and sharing the benchmark with another analyst or builder.
Turn filings, calls, peer metrics, and market reactions into a citation-bound investment thesis that can also say: no convincing thesis yet.
Try the live demo | Quickstart | Methodology | Roadmap | Contribute
Watch the 91-second narrated AAPL walkthrough | Source event -> causal bridge -> peer evidence -> reverse DCF -> Bull/Bear/Judge
Most AI research tools are good at producing text. IC Copilot is designed to produce an auditable investment decision trail:
- Exact evidence: filing section, period, metric, unit, source tier, URL, and parser status.
- Causal reasoning: source event -> business driver -> KPI -> earnings/FCF -> valuation -> catalyst.
- Adversarial review: Bull, Bear, and Judge views with unresolved contradictions kept visible.
- Market-implied expectations: reverse DCF and implied operating assumptions when historical consensus is unavailable.
- A real ability to say no: boilerplate, stale facts, unmapped drivers, and unsupported LLM claims do not become high-conviction ideas.
The app is built for equity analysts, credit analysts, portfolio managers, and AI builders who want evidence before eloquence.
- Watch the narrated AAPL walkthrough, or open the hosted app. The AAPL Deep Initiation demo is loaded automatically.
- Read the one-screen IC Story: verdict, what changed, causal bridge, counter-thesis, valuation, and next action.
- Open an evidence drawer to inspect the exact citation and formula trace.
- Switch to NVDA, ORCL, BABA, TSLA, or GS to see a different research playbook.
The frozen demos require no API keys, make no paid provider calls, and contain no licensed full-text payloads.
flowchart LR
A[Entity] --> B[Sources]
B --> C[Claims]
C --> D[Business drivers]
D --> E[Peer read-through]
E --> F[Company model]
F --> G[Valuation]
G --> H[Price attribution]
H --> I[IC thesis]
I --> J[Monitor and post-mortem]
Every connection is scored separately. A weak link identifies the exact missing evidence instead of turning into generic uncertainty.
| Demo | Research lesson |
|---|---|
| AAPL | Source-backed cash-flow and margin thesis with model-derived scenarios |
| NVDA | Neutral-first capex and goodwill investigation with semiconductor peer checks |
| ORCL | Cloud/RPO growth versus data-center capex, financing, dilution, and cash conversion |
| BABA | ADR/FPI normalization, China segments, bilingual external context, and price-only capture |
| TSLA | Driver-specific peer metric read-through across automakers |
| GS | Bank/broker economics, equity lens, and credit lens |
| SPCX | Entity-resolution guardrail that stops unreliable analysis early |
git clone https://github.com/mwsp-code/IC-Copilot.git
cd IC-Copilot
python -m pip install -r requirements.txt
streamlit run app.pypython server.py --port 8501The lightweight server automatically tries ports through 8510 when 8501 is occupied.
python -m pip install -e ".[dev]"
pytest -qdocker build -t ic-copilot .
docker run --rm -p 8501:8501 ic-copilot- Fast Screening: 4 quarters, 2 annual reports, 4 calls, and the highest-ranked anomaly.
- Adaptive IC Research: default; 12 quarters, 4 annual reports, 12 calls, and five material changes.
- Deep Initiation: 20 quarters, 5 annual reports, 20 calls, and deeper contradiction work.
- Investigate This Event: a filing-, call-, metric-, or news-scoped workflow.
Ambiguous changes are neutral first. Constructive and adverse explanations are investigated side by side before direction is assigned.
SEC and issuer evidence remain the backbone. Official macro, market data, paid providers, Wisburg, and news can deepen context, but source tiers and licensing policies remain visible.
LLMs may plan registered sources, triage documents, draft structured extractions, compare evidence, and write an IC narrative. They cannot invent citations, targets, calibrated probabilities, or promotion decisions. Deterministic validation remains authoritative.
API keys are optional for the demos. For live work, save credentials through the OS keychain, environment variables, or Streamlit secrets. Never commit keys or licensed payloads. See Provider Configuration.
The repository ships a no-network benchmark over 25 research-integrity cases across AAPL, NVDA, BABA, TSLA, and GS. It checks event grounding, thesis grounding, counter-thesis handling, monitor/payoff readiness, and promotion integrity.
python scripts/run_benchmark.pyThis is an integrity benchmark, not a claim of investment performance. See Benchmark Methodology.
IC Copilot separates deterministic evidence collection from hypothesis generation, valuation, monitoring, and optional LLM synthesis. Provider failures are isolated, normalized, and source-labelled; missing values remain Unknown rather than becoming zero.
See Architecture and Research Methodology.
Useful first contributions include:
- Add a sector KPI playbook.
- Add an ADR/FPI profile.
- Add a canonical metric alias.
- Add a registered source adapter.
- Add a sanitized demo or benchmark case.
The UI and docs contain explicit contributor paths. Read CONTRIBUTING.md, browse issues labelled good first issue, or start a design question in GitHub Discussions after it is enabled.
IC Copilot is research software under active development. It does not execute trades and does not provide financial advice. Frozen demos illustrate methodology; refresh live sources before using any observation in an investment process.
See Roadmap, Security Policy, and Changelog.
If IC Copilot improves how you think about evidence, consider starring the repository and sharing the benchmark with another analyst or builder.