TK, this one's yours.

New direction for HookPilot: an AI-driven analytics product for non-profits. HookPilot doesn't just show a client their donor data, it runs a deep analysis and hands the Executive Director a ready-made plan — donor optimization opportunities, ad spend reallocation (flagging underperforming Facebook ad spend and suggesting UGC instead), and specific next steps, not generic charts.

What exists right now: the domain and this frontend, hosted on Base44. No backend data pipeline, no dashboard, no AI analysis layer yet. That's what this page covers.

A data warehouse plus embedded BI dashboard, self-hosted, not a paid vendor. Going with self-hosted Apache Superset (open source, same engine Preset's paid product is built on) instead of a SaaS embedded-analytics vendor like Luzmo. Reasoning: near-zero budget, AWS infrastructure already available (from Lurexo/previous clients), and Superset supports embedded dashboards natively via guest tokens so it can be dropped into this Base44 frontend via iframe — no vendor license fee, just hosting cost and setup time.

Concrete steps:

• Stand up a Postgres database on the AWS infra. This becomes the single source of truth both the dashboard and the AI analysis layer read and write from.
• Deploy Superset via Docker (docker-compose for a first pass, ECS later if this grows past MVP).
• Configure Superset's embedded/guest-token feature so a dashboard can be securely iframed into this site, scoped to one client's data at a time.
• Scope to a single client for now (likely VWB). No multi-tenant isolation yet — that's real additional work for once there's more than one paying client, flagged as a known gap, not something to solve now.

A backend analysis job that calls the Claude API. This is the real product, not the dashboard. A backend service (Python or Node, TK's call) that ingests a client's donor CSV export into the Postgres database, calls the Claude API (Anthropic's Messages API) with analysis instructions Femi will write — the specific things HookPilot looks for: donor lapse risk, giving pattern segmentation, mid/major/legacy gift candidates (using ZIP3, the first three digits of the ZIP code, joined against public Census income data, not the full ZIP), and eventually ad spend efficiency once Meta Ads integration exists. Writes the output — scores, flags, a narrative summary — back into Postgres as new tables Superset can visualize.

Femi owns writing the actual analysis instructions and prompts; TK builds the pipeline that gets data in, calls the API, and gets results back into the database in a shape Superset can chart.

Bare-bones demo-ready (Postgres and Superset deployed, embedded into this site, one sample dataset loaded and charted, no AI layer yet): roughly 3–5 focused working days, assuming standard Docker/AWS setup, no major surprises.

With the Claude API analysis pipeline working end to end (CSV in, real scored output out, charted): add another 3–7 working days, depends heavily on how much iteration the analysis logic needs once real output is visible.

These are estimates, not commitments. Confirm against actual current bandwidth. If the honest answer is "can't get to this soon," better to know now than have it slip silently.