Encode your thinking once.
Let the agent execute on it forever after.

Nagini is an operating system for buy-side analysts to encode their thinking once and let an AI agent execute on it forever after. Not a co-pilot. The analyst's own framework, applied at machine speed, with cell-by-cell audit. Modern LLMs are good enough at financial reasoning that the bottleneck is no longer “can the AI do the math?” — it's “does the AI think the way this analyst thinks?” Markets are made of differing perspectives; an agent that operates on a generic framework is an undifferentiated agent.

Pre-revenue, pre-customer — but the V1 surface is real. As of 2026-05-07:

• A ten-scenario live walkthrough passes 10/10 against production: magic-link sign-in, real-ticker thesis creation, streaming agent reply with read_cell, a real run_dcf round-trip to a Python worker on Cloud Run, and write_cell with durable upsert, audit row, DOM re-render, and DB-confirmed value flip.
• Two-tenant isolation enforced via Postgres RLS plus the withTenant middleware plus JWT-bridged scope (ADR-0004; ADR-0009 closes the cross-tenant signin escape hatch).
• Multi-LLM agent loop. Gemini 2.5 Pro default, switchable per turn to Claude Opus 4.7.
• 9 ADRs locked, and counting. 56 PRs landed since project start (2026-04-29) — including a five-PR burst from a parallel-agent team in 30 minutes.
• 93 web tests + 64 worker tests + the ten-scenario live walkthrough; no production incidents.

Mohammad Khodadadi · prior background in TBD · what I'm betting: TBD.

Nagini ships with a team of agents, not a team of humans. One founder plus a fan-out of Claude and Gemini agents — each scoped to a single milestone, each merging through the same review gate as a human PR. An orchestrator agent schedules the fan-out, reads the coordination ledger, and prevents two agents from racing on the same file. The live ledger and decision log are public. The pattern is documented in ADR-0007 and ADR-0013 — most products don't ship their architecture this transparently.

Every binding architectural and product call is written down. Read the rationale — approach selection, cell architecture, tool registry, auth + RLS, the design system, deploy topology, the team-of-agents pattern, the cross-tenant signin escape hatch, the M6 data-ingestion source picks, the V2/V3 wedges — under public ADRs at docs/decisions. If the rationale matters to you, that's where it is.