Back to selected workCase 02 / Government-affairs strategy · System design
AI that uses a firm’s specialist knowledge to build strategy.
I designed an AI system to create strategies and make decisions using knowledge held by experienced government-affairs practitioners, the firm’s growing institutional memory, and changing state and legislative timelines.
- My role
- Product owner and lead designer: strategy-generation model, UX, AI direction, and engineering specifications
- Context
- Government-affairs platform
- Delivery
- Designed; implemented by engineering
01 / The problem
Make expert strategy available inside the product
The task was to design a system that could generate a relevant strategy and make decisions within it. The strategic knowledge it needed lived with a small group of experienced government-affairs practitioners. There was no suitable public dataset of vetted strategies in the required format. The firm’s expertise and accumulated institutional knowledge had to inform the reasoning.
That strategy also had to remain aligned with fluid state and legislative timelines and overlapping budget cycles. A useful recommendation depends on the firm’s context and the timing of the opportunity, both of which can change.
02 / My decisions
The choices behind the product
01Collect knowledge that public sources could not supply
The design needed a way to use a small group of practitioners’ non-public judgment in strategy generation and decision-making. With no suitable public dataset available, I built an authoring instrument and standards to collect their vetted strategies in a format a model could later use. Public information could provide context, while practitioners supplied the expert examples. Collection was still underway at handoff; I did not treat a fine-tuned model as an accomplished result.
Use institutional memory to make strategy relevant
I designed strategy generation around the firm’s accumulated knowledge, rather than treating each client goal as an isolated prompt. Institutional memory was intended to compound as the firm used the product, giving subsequent strategies more relevant context for their decisions.
Align decisions with changing timelines
State and legislative calendars are fluid, and budget cycles overlap. I designed the system to ground strategy in those timelines, check sequencing, and monitor changes so it could propose updates when the timing or dependencies shifted.
Check the strategy before showing it
I specified deterministic checks for calendars, budget rules, and step dependencies before a draft reaches the user. A failed draft should be corrected or held. These checks address the parts that can be verified mechanically; they do not establish that every strategic judgment is good.
Keep proposed changes tied to the current strategy
The monitoring agent proposes changes on the strategy itself. Suggestions belong to a specific version. Accepting one requires re-checking and refreshing the remaining suggestions against the updated plan. I diagnosed two logic gaps and specified their fixes.
Make missing context actionable
When the system lacks context, it should try to resolve it. If it cannot, it should name the missing information and offer a clear way forward rather than present an unsupported strategy.
03 / Delivery
What I delivered
I delivered a functional prototype and a twelve-point engineering requirements set covering the strategy model, grounding, checking, human review, and actionable steps. Following CTO and founder review, engineering implemented the feature in production. I do not have measured adoption or model-quality outcomes to report.
04 / The open question
What I’d test next
I would establish an evaluation loop earlier. The key question is whether the verifier catches genuine errors without suppressing useful, unconventional strategies. Correction rates and error categories would help set that boundary.
About the evidence
Based on my anonymized AI portfolio and contribution record. My contribution was product ownership, design, AI direction, and engineering handoff; engineering implemented the production feature.