I built Worldwide Planning Intelligence as a research prototype on local hardware. It is not a hosted product. This page covers what I built, how I tested it, and what I found. Full write-up: white paper (Word). Short overview: tool page.
When headcount doubles in a year, the planning stack built for a smaller org stops working. Risk ends up in spreadsheets. Dependencies get tracked by hand. Each program ends up with its own version of the plan, and the TPM spends time maintaining the register instead of running the program.
The prototype includes a tiered risk register (P0–P3), guided intake that turns plain-language answers into scored records, a dependency log for capital and cross-program conflicts, an executive weekly report in pyramid format, and a benchmark runner that scores local LLMs on the same triage prompt. Stack: Python, Streamlit, Ollama.
To reproduce the benchmark, build a register of 50–100 risks, write one triage system prompt, define a rubric (tier accuracy, field completeness, mitigation quality), run every risk through each model, and compare accuracy against latency. I ran 82 risks across five local models.
Results:
- Llama 3.1 8B: 76% rubric score, 7.5 sec/call
- Gemma 3 27B: 73%, 46.7 sec/call
- Mistral Small 24B: 66%, 26.2 sec/call
- Qwen 2.5 32B: 56%, 38.6 sec/call
I would pick Llama 3.1 for production triage on local hardware.
Cloud API cost was about $0.009 per call, roughly $90/year at weekly volume. Local inference was $0 per call on hardware I already owned. Cost matters, but privacy mattered more. Every cloud call sends program data off-premises.
A few things held up in testing. Smaller 8B models beat larger 32B models on speed-adjusted triage quality. RAG with historical decisions improved grounding in manual review. Guided intake cut risk creation from about 15 minutes to about 3. One register with tier filtering replaced three separate tracking documents in the workflow.
This is not a hosted product. If you want to walk through the setup, reach out. For tools you can run in the browser today, see the Program Risk Register and the tools catalog.