Case study / 01
LegiKit
AI drafting built for congressional work — vote recommendations, briefing memos, and constituent letters grounded in live bill data and office voice. 30+ staff users.
Phase / 01 Discover
Map the workflow.
I was working as a Legislative Assistant in the Senate when I noticed the same pattern across every office: staff drafting vote recs, briefing memos, and constituent letters by copy-pasting bill text out of Congress.gov into blank Word docs, with the member's positions and writing style living entirely in the heads of whoever had been there longest. ChatGPT couldn't fix this — it didn't know the office, the bill, or the format. I mapped the workflow by doing it myself, then watching how legislative correspondents, LAs, and LDs actually moved a document from "we need a response" to "ready for the chief's signature."
Phase / 02 Build
Ship the system.
LegiKit is a workflow tool, not a chatbot. The frontend (React + TypeScript) routes you by document type — vote rec, briefing memo, constituent letter, talking points — so the system structure matches what the office already uses. The backend (FastAPI, Supabase Postgres, LLM APIs) pulls live bill data from a Congress.gov pipeline I'd built, layers in the office's prior positions and house style, and streams a Word-exportable draft. The hardest design choice was resisting the temptation to make it general-purpose — workflow specificity was the entire wedge.
Phase / 03 Deploy
Earn trust.
Procurement on the Hill kills most software before it reaches a user, so I bypassed it: direct onboarding, in-person walkthroughs, building trust one office at a time. Every new office got a short setup conversation where I loaded their position anchors and exemplars by hand before walking them through their first generation. That hand-rolled distribution earned the trust institutional sales never could — 30+ staff across multiple offices now use it.
Phase / 04 Compound
Reuse the learning.
Each office I onboarded sharpened the product in two directions: better defaults for the next office, and richer office memory for the current one. The Congress.gov pipeline got more reliable because users were generating docs live during hearings. New document types got added because staff asked for them. The biggest compounding effect was institutional — when a staffer leaves, the system still knows what the member has said about housing, taxes, immigration, and how the office writes about each.