Prashanth

Anubhav is a team of AI agents on Meta Muse that learns the rules a business never wrote down, from its own records, and plans better because of it. The problem: small Indian truck fleets plan on phone calls and memory. A dispatcher knows a 32-foot truck can't enter one godown in Jangaon. Three trucks were turned back, but no software knows it. Asked for a truck, a model passes every written check and still picks the wrong one. What Anubhav does: a Lesson Miner reads 90 days of trips and drivers' notes and proposes unwritten rules. An evidence gate in code tests each one against the records (it holds, is rejected as a habit, or is stored as expired), and a person approves it. The same model, asked the same question in Roman Telugu, then picks a 22-foot truck and names the rule it followed. It only ever drafts: a person dispatches. A Planner leads a team of an Analyst, Tool Forge (read-only SQL that can become reusable tools) and a Lesson Miner. Each agent can call only its own tools, enforced in code, and none can change the records. Approved regional facts can be shared, so a new company starts with them on day 1. Results from 21 pre-registered experiments, failures included: unwritten-rule cases went from 48% to 90% correct on Muse Spark Open and 45% to 90% on Muse Spark 1.3, with no unsafe drafts; a new company went from 33% to 100% on day 1. Data is synthetic but grounded in public sources. Built solo during the hackathon; every commit is public. Live: https://huggingface.co/spaces/KPrashanth/anubhav Code: https://github.com/prashanth-ds-ml/anubhav