Project MIT
live
Global Operations Risk Monitor
How do you monitor operational risk across a supply chain?
Operations Risk Monitor is a multi-agent framework that watches operational risk across freight, fuel, labour, weather and geopolitics, then turns what it finds into briefings you can actually use. Seven specialist agents collect and filter the feeds into one weekly briefing.
Operational risk hides in a hundred feeds at once
Ocean freight, fuel prices, port strikes, weather, sanctions, congestion - the signals that move a supply chain are scattered across news feeds, carrier advisories and authority sites. By the time someone has read enough to see the pattern, the window to act has often closed.
One person can't watch all of it. And a single general-purpose summary flattens the detail that matters - which carrier, which lane, which region.
Seven specialist agents, one briefing
The Risk Monitor runs a fleet of specialist agents, each owning one slice of risk - freight, fuel, diesel, strikes, weather, geopolitics and port congestion. Each watches its own sources and writes its own briefing, so the detail survives instead of being averaged away.
It runs two ways: a CLI that produces weekly reports (HTML, Markdown, PDF and slide decks) and a live dashboard that refreshes itself every few hours. The model behind the summaries is yours - Anthropic, OpenAI or a local Ollama model - switched with a single line of config.
Everything the model writes is marked as AI-generated - a line you can read, and metadata a machine can detect, in every format it exports. Text lifted straight from a source is left unmarked, because labelling it would misstate where it came from. That's Article 50 of the EU AI Act handled in the output itself rather than in a policy page.
Collect, filter, summarize, compose, save
One base class defines the five-step contract; a registry wires each agent in. No orchestration framework underneath - subclass the base agent, register it, done.
Specialist depth, model-agnostic, config-driven
Seven agents each own a slice of risk, so a briefing keeps the detail - which carrier, which lane, which region - instead of flattening it into one vague summary.
Everything routes through LiteLLM. The same code runs on Anthropic, OpenAI or a local Ollama model. Change one line of config and set the key - no rewrite.
A CLI for deep weekly reports - HTML, Markdown, PDF and slide decks - and a self-refreshing dashboard for the live view, sharing the same agents.
Regions, carriers and sources live in YAML; edit the watchlist without touching the pipeline. Add a new agent by subclassing one base class and registering it.
Where the Risk Monitor is heading
Bring freight cost analysis into the fleet - the one slice of risk still stubbed out - so spend sits alongside disruption in the same briefing.
More regions and carriers, added through config alone, so the same harness scales to new trade lanes without new code.