in development
Reconstructing Wealth
AI prepares, the advisor delivers.
Financial advisors are in the midst of a seismic shift. An extraordinary wealth transfer. A consolidated stock market. Millennials who want proof before they trust. And AI that can do part of their job for free.
They need to do the human parts of the job at scale, and establish their own investment edge. So we built a platform for exactly that. AI preps the back office — meeting prep, follow-ups, earnings and filing summaries, scanning for estimate revisions. The advisor delivers every word. AI never sends a client message or recommends a trade, and the compliance rules live in the database, not in good intentions.
At a glance
A solo or small fiduciary RIA runs on a fragmented, expensive stack — CRM, spreadsheets, a research analyst — under strict SEC books-and-records obligations that generic SaaS doesn’t satisfy.
I built a monorepo of three apps over one compliant data core: marketing and lead flow, an advisor CRM with a read-only client portal, and an equity-research engine — bound by a multi-provider AI router and event-driven workflows that prepare the work for the advisor to deliver.
Bounding AI inside a fiduciary context meant encoding “AI prepares, the advisor delivers” everywhere — no client-facing sends, visible authorship, review-before-commit — and implementing the SEC books-and-records controls (append-only audit, soft deletes, access logging) purely at the Postgres and RLS layer.
It’s in development, and already shows a single advisor running lead-gen, CRM, and institutional-style research from one system for a few hundred dollars a month — with SEC books-and-records controls built in from day one.
AI components
A provider-agnostic router dispatching tasks across Claude (Opus/Sonnet), GPT-4o, and Gemini with per-task primary/fallback chains and cost and latency budgets.
Intentionally not autonomous — deterministic multi-step enrichment pipelines and a signal→CRM bridge rather than open-ended agents, because the fiduciary answer is always human-in-the-loop.
Structured/relational retrieval — each task is handed exactly the records it needs from Postgres, plus on-demand long-document fetch; no vector store.
None yet — full cost/latency/success telemetry per task is in place as the foundation an eval harness would build on.
18 event- and cron-driven Inngest functions (meeting prep, earnings and filing summarization, EPS-revision signal detection, quarterly briefs).
Scheduled research jobs across FMP, Finnhub, and SEC EDGAR with dedup-before-insert idempotency and graceful degradation.
Technology
Next.js • Supabase • Inngest • Claude • GPT • Gemini