NexussAI

One hundred litres of diesel. One blocked road. Two hundred people. Four hours until the hospital generator dies. THE PROBLEM Kottur is a panchayat town at the foot of the Western Ghats, on the Kerala border — the town I know firsthand. On 22 October 2026 a monsoon landslide severed the Pollachi-Valparai, India corridor on NH-544 and kept it shut for 15 hours. The fuel depot was cut off, the hospital generator had four hours of diesel left, the water pump was dead, and 200 people across three shelters had no water. Nobody was short of information that day. They were short of a decision. WHAT I BUILT LIFELINE AI is an agent that makes that decision and then defends it. It treats the town as a body: diesel is blood, the hospital is the heart, the water pump is the lungs, the shelters are the organs. That is a dependency graph, not a metaphor, and the agent sorts it topologically before it spends a litre. Eight steps run per incident: perceive the flood, listen to a field report, order the dependencies, route around the blockade, allocate the reserve, check the budget, ask a human once, replan when the world changes. Seven of the eight need no human. The single gate that remains is code, and its default is no. THE MOMENT THAT MATTERS (video, 1:51) The model proposes. Code disposes. In the recorded run a human operator insists the pump outranks the hospital. The agent refuses, names both violations — the heart would be starved, the total would exceed the reserve — and applies the 70 L / 30 L split anyway. It accepts a changed world and refuses a changed priority. Obedience is the easy default; this one is built to be argued with. SIX META MODELS, SIX DIFFERENT JOBS SAM 3.1 — measures flood extent per frame from real pixels, not a caption. Muse Voice Transcribe — turns a Tamil field call into text. Muse Wonder — second opinion, and reviewer of every generated payload. Muse Spark — drafts the SMS, spoken alert and decision sentence in eight locales. Muse Spark Open (muse-spark-open-1026, Baseten) — proposes the allocation itself. MMS-TTS (open weights, this laptop's CPU) — speaks the alerts at 0.23x real time. WHY IT HOLDS UP The deterministic core is Python standard library only: Kahn's algorithm, A* with a haversine heuristic, a constrained 70/30 solver, a hash-chained ledger. Median route search 0.027 ms, allocation 0.014 ms, validator 0.005 ms. A* is audited for admissibility and checked against brute-force enumeration. Every Spark Open allocation is re-validated by the same code that rejects a bad human one. 110 acceptance tests. python judge.py prints ten checks with the exact command beside each. 60 counterfactuals, each a real solver run. Ledger at $23.42 of $150. No API key needed; LIFELINE_OFFLINE=1 runs the core with the network cut. SCALE Nothing in the code knows the word "Kottur". Organs, roads, reserve and demands are data in town.json, and 26 towns across 9 regions run on the same three algorithms. Only Kottur has a surveyed fuel scenario; every other pin returns a labelled template rather than an invented litre count. 1.81 billion people are directly exposed to 1-in-100-year floods (Rentschler et al., 2022). Kottur is the proof; the pattern is the product.