Sky Solutions · 30 Apps in 30 Days

Building an evidence-first AI research platform with researchers in control

I independently designed and directed the AI-assisted development of Sky AI Research (Sky AIR), a qualitative research platform that keeps source evidence, provenance, and researcher review visible throughout analysis. I defined the product strategy, UX architecture, design system, implementation specifications, and quality controls for a completed first MVP that remains under development.

UX research
Product design
Responsible AI

Sky AIR evidence-first qualitative research platform

Functional MVP completed

Evidence workflow stages

Full stack

React, FastAPI, and PostgreSQL

Human-reviewed

Design and implementation handoffs

Project overview

Two questions shaped the project

Product question

Could AI make research knowledge easier to access without separating conclusions from their evidence? Business analysts and product teams regularly relied on UX researchers for answers, making the UX team an information bottleneck.

Delivery question

Could AI accelerate an end-to-end design and development pipeline without sacrificing consistency, accessibility, or engineering control? I independently defined the product and specifications; Codex generated implementation code under my direction, and I reviewed, tested, and iterated the first MVP.

Part 1 · Designing the product

Creating trusted self-service access to research knowledge

Sky AIR serves three connected audiences while keeping reviewed evidence—and the UX team’s research standards—at the center of every answer.

Three audiences and two shared safeguards

UX researchers and designers · review findings and maintain knowledge
Business analysts · explore reviewed evidence independently
Product team · revisit user needs and decision rationale
Shared value · faster answers grounded in inspectable evidence
Shared boundary · AI access does not replace UX interpretation

Routine questions become easier to answer while UX researchers and designers preserve more time for new research.

Sky AIR intended users and self-service research workflow — Researchers maintain reviewed knowledge while business analysts and product teams gain trusted self-service access.

Sky AIR self-service question and evidence workflow — Business analysts and product teams can ask routine questions while researchers retain control of reviewed knowledge.

Product strategy

Designing for evidence before automation

The product strategy centered on making AI useful without asking teams to trust opaque outputs. Every finding and answer needed a visible path back to source evidence.

Risk

Fast AI summaries without traceability

Conclusions can become detached from source context
Teams may confuse generated language with reviewed findings
Research rationale can disappear as work moves forward

Direction

Evidence-first assistance with human control

Keep citations and provenance visible
Separate generated, reviewed, and approved states
Let humans correct findings before knowledge is reused

Sky AIR human oversight model from evidence through reusable knowledge — Trust is represented as a workflow with visible evidence and explicit review states.

Information architecture

Structuring research as an inspectable evidence system

I designed connected research objects so transcripts, codes, findings, reports, reusable knowledge, and questions could retain provenance as they moved through the product.

Sky AIR evidence-system information architecture — Connected research objects retain citations, provenance, ownership, and review status.

5Connected workflow stages

3Visible review states

Evidence citations, provenance, ownership, and review status make it possible to inspect how knowledge was created before it is reused.

Core workflow

From transcript to evidence-grounded answers

The five-stage workflow turns raw research into reusable knowledge without hiding the evidence chain or the researcher’s review decisions.

Five stages, four traceable handoffs

Transcript → AI-assisted coding with evidence spans
Coding → researcher-reviewed session report
Session report → approved record knowledge
Record knowledge → cited answers in Ask & Explore

At each handoff, users can inspect evidence, understand status, and correct the material before it becomes organizational knowledge.

Five-stage Sky AIR evidence workflow — The end-to-end evidence workflow keeps researcher review visible at every handoff.

Responsible AI

Making AI assistance visible, reviewable, and correctable

Responsible AI was designed into the workflow and architecture—not added as a disclaimer. The interface shows where content came from, what the AI generated, and what a human has reviewed.

Evidence + provenance

Coded finding with citation and source context — Citations, source spans, ownership, and retained evidence

Human review + correction

Generated, reviewed, and approved states — Researchers verify, edit, approve, or reject before reuse

Safe, swappable architecture

Provider boundary and deterministic test path — Synthetic/public data, no persisted raw keys, and explicit insufficient-context behavior

Part 2 · Directing the AI-assisted build

Directing an AI-assisted design-to-delivery pipeline

I defined the strategy and specifications, then directed AI-assisted implementation across seven connected stages. I manually reviewed behavior, consistency, accessibility, and test evidence before approving each handoff.

ChatGPTProduct strategy, architecture, UX critique, and specifications
Figma + design systemFlows, states, reusable tokens, and patterns
Storybook + ReactDirected AI-assisted component creation and reviewed behavior, states, consistency, and accessibility
FastAPI/PostgreSQL + testingApplication programming interfaces (APIs), provenance, automated tests, and Playwright walkthroughs

ChatGPT → Figma → Design system → Storybook → React → FastAPI/PostgreSQL → Testing
Human review and approval at every handoff

Sky AIR AI-assisted design-to-delivery technology stack — The full delivery stack connected product strategy and interface design to implementation, testing, and human approval.

Validation + outcome

A functional MVP with live OpenAI integration

The first MVP is complete and functional. It includes a reusable Figma design system, an AI-assisted Storybook component repository, React front end, FastAPI/PostgreSQL back end, live OpenAI integration, automated front- and back-end tests, Playwright acceptance walkthroughs, and human review controls. Additional vertical slices and formal user evaluation remain underway.

Sky AIR Storybook and implemented component evidence — The first vertical slice connects research intake through cited answers.

Sky AIR current product status and next steps — Generated, reviewed, and approved states make knowledge status explicit.

Functional product

A working full-stack application connects transcript intake through evidence-grounded answers.

Review and test controls

Automated tests, Playwright walkthroughs, and human approval support each handoff.

Ongoing development

Additional vertical slices and formal target-user evaluation remain underway.

Designing AI responsibly meant treating trust as a workflow—not as a message added after the fact.

Reflection · Sky AIR MVP

Contact

Need a researcher-designer for a complex public-sector system?

Let’s talk about enterprise modernization, accessible workflows, or responsible AI.

snaggums@gmail.com