AutoInvestigate

Folder	What	Run
website/frontend/	the three-column app (no build step)	served by the backend at http://localhost:8090/, or website/frontend/run.sh
website/	Part 2, idea evaluation: 13 agents on AGNTCY/SLIM, models via OpenRouter	cd website && ./setup.sh (see website/README.md)
artifactfactory/	Part 3, idea execution: turns an idea into artifacts (the frontend's column 3)	./run.sh
evalreport/	report viewer	see the folder

Quick start for the evaluation backend: install Docker, get an OpenRouter key, then cd website && ./setup.sh and open http://localhost:8090/.

---

Outshift Research Intelligence

A runnable Streamlit MVP for specialist emerging-technology analysis with Business, Venture Capital, and Technical perspectives, an independent critic, and a general-purpose baseline comparison.

Requirements

• Python 3.10+
• An OpenRouter API key with access to the model you select

Run everything with Docker Compose

A single docker-compose.yml at the repo root builds and starts every service:

Service	Host URL	What it is
frontend	http://localhost:8000	Static three-column workspace (no build step)
artifactfactory	http://localhost:8080	Artifact Factory REST API (idea → plan → artifacts) — column 3
web (+ slim, agent_host)	http://localhost:8090	Evaluation backend (multi-agent deliberation) — column 2
evalreport	http://localhost:8501	Streamlit "Outshift Research Intelligence" app

cp .env.example .env      # then set OPENROUTER_API_KEY (and OPENAI_API_KEY for evalreport)
docker compose up --build

Everything runs on a single OpenRouter key. The root .env holds one OPENROUTER_API_KEY, and every service reads that one file (website/ and evalreport/). The evaluation backend also needs a SLIM_SHARED_SECRET; cd website && ./setup.sh writes both for you.

Then open http://localhost:8000 (frontend). The evaluation backend also serves the same frontend at http://localhost:8090. Generated artifacts and daemon state persist to ./_artifacts and ./_artifactfactory-data on the host. Stop with docker compose down.

Each service also has its own Dockerfile (artifactfactory/, website/frontend/, website/, evalreport/) if you want to build or run one on its own.

Permission errors

Docker creates a bind-mount source directory as root when it does not already exist, so a container running as you cannot write to it. You then see errors like PermissionError: [Errno 13] Permission denied: 'logs/runs' or sqlite3.OperationalError: unable to open database file. The services fix this on startup — their entrypoint chowns the mounted dirs (logs, website/data, website/council_results, _artifacts, _artifactfactory-data, _ideation-data) to HOST_UID:HOST_GID and then drops to that user — so a rebuild repairs it:

docker compose up -d --build

To repair the directories without rebuilding, run the helper script:

./fix-permissions.sh            # create + own every mounted directory
./fix-permissions.sh --restart  # ...and rebuild + restart the services

Run on macOS/Linux

cd outshift_research_intelligence
python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements.txt
# OPENROUTER_API_KEY is read from the root .env (../.env); or export it here:
export OPENROUTER_API_KEY="sk-or-..."
streamlit run app.py

If streamlit: command not found, use:

python -m streamlit run app.py

How it works

• Business evaluator: customer pain, market, willingness to pay, strategic fit, adoption friction (30% total decision weight).
• Venture Capital evaluator: upside, timing, moat, capital efficiency, growth, risk-adjusted return (25%).
• Technical evaluator: feasibility, architecture, performance, security, prototype speed, operations (45%).
• Independent critic: challenges assumptions and proposes a measurable next experiment.
• Baseline: a single general-purpose analysis of the same research objective.
• Export: specialist report as JSON and comparison as CSV.

Important limitations

This is a specialist agent workflow using an existing language model API, not a newly trained foundation model. It does not guarantee better performance than ChatGPT. The built-in comparison dimensions are only rough heuristics and are not an objective benchmark. To substantiate a performance claim, evaluate both systems on a fixed test set with blind human scoring, factual verification, and repeated trials. The current MVP does not perform web retrieval or verify citations automatically.