SLR.AI · AI / B2B SaaS · Government
AI screening platform for systematic literature review, built for US government agencies and commissioned by PD4 Solutions.
The work covered setup, bulk screening, manual review with the AI Scoring Panel, a modular dark-theme design system, responsive layouts, and the user flow for review creation and reference handling. I designed the platform as sole Product Designer, in close cooperation with the Product Manager. The prototype advanced into development.
The AI Scoring Panel was the load-bearing decision. AI reasoning was placed on the same screen as the verdict, in four layers stacked together.
The AI Composite Score is built from two components shown explicitly: an NLP score and a Historical score, each with its weighting percentage on the surface. The weighting is configurable rather than fixed.
Beneath the scores, the Match Criteria summary lists how many inclusion and exclusion criteria the AI matched, with a short evidence excerpt directly underneath each criterion. The same screen also shows the abstract of the paper itself, with the sentences the AI used as evidence highlighted in place. A custom question field sits below, taking free-text questions answerable from the paper view.
The four layers stacked together let federal users verify the AI's reasoning at multiple depths without leaving the screen. Score, criteria, source, and free-form question are four different angles on the same verdict.
US government agencies use systematic literature review to ground research that informs publications and public statements. The work is exhaustive by design and the reasoning has to be defensible to outside scrutiny. PD4 Solutions framed the engagement as compressing review cycles that had taken over a year into a couple of months by introducing AI screening across the corpus.
Speed at that compression rate is only useful if the audience can verify the AI's reasoning. A verdict reached in months rather than a year is rejected by federal users if its basis cannot be examined. The brief made this explicit: ensure that users understand and trust results of AI screening.
Discovery interviews with domain experts informed the framing. Trust at this scale is not a function of the AI's accuracy claim. Trust is a function of whether the user can interrogate the verdict.
Around the AI Scoring Panel, the platform was delivered as six functional surfaces, plus the design system that built them and the user flow that connected them.
A three-step intake covering review naming and type, data source filtering by date range and language, and topic-level inclusion and exclusion criteria. The setup determines what the AI screens and what it screens for.
The Ranker screen carries an All Scores summary alongside three score buckets: HIGH 8-10, MEDIUM 4-7, LOW 0-3. Bulk review treats each bucket as a unit. The user can apply Yes, No, or Maybe across the bucket, or step into individual papers from the same screen.
When the AI's verdict needs examining, manual review opens the AI Scoring Panel against a single paper. Yes, No, and Maybe verdict buttons sit at the bottom of the screen. Reference-by-reference navigation moves through the queue.
A modular dark-theme component library built for iteration speed. Nine visible categories: dropdowns, text inputs with chip pattern, buttons in multiple states, radio buttons, checkbox lists, score cards, icon strip, tooltip, toast and notification variants.
Two viewport sizes delivered. The Import References intake and the Bulk Review queue both adapt without losing the criteria-and-evidence pairing that anchors the screening logic.
A sitemap-style diagram covering review creation through duplicate handling outcomes. Decision points and their resolutions are colour-coded, including the duplicate confirmation path that ends in either retention or permanent deletion.
AI screening that could be questioned. Built for federal research environments, advanced into development.