Veritone · AI Investigative Platform
Helping investigators understand what the AI found—and what to review next
I designed enterprise workflows that made it easier to search complex evidence, interpret confidence scores, and keep investigators in control of consequential decisions.
Explore the case studyPortfolio reconstruction of product workflows; sensitive data omitted.
✦Key insightCross-source match establishes a sequence of events.
Candidate matchHigh confidence
92%Location confirmedHigh confidence
89%Document verifiedHigh confidence
94%Audio matchHigh confidence
91%✦
Coherent storyA timeline constructed from reviewed, correlated evidence.
The opportunity
AI could find possible matches quickly. Investigators still needed to understand why those matches mattered.
AI can surface a promising moment in seconds. The product still has to help a person understand why it matters—and decide what to do with it.
As Lead UI/UX Designer, I owned end-to-end design across discovery, workflow analysis, prototyping, validation, and delivery for enterprise investigative products.
The central design challenge was making complex evidence and machine-generated matches understandable enough for an investigator to review, question, and act on responsibly.
- Users
- Public safety, legal, and compliance teams
- Core material
- Video, audio, documents, images, and metadata
- My ownership
- Strategy through validated interaction design
01
Challenge
Investigators had plenty of evidence. The hard part was understanding how it connected.
Evidence could arrive as
VIDEOAUDIODOCUMENTSIMAGESMETADATA
Different formats.Different sources.
One investigation.
Discovery was ambiguous
Investigators needed to find relevant moments across large, mixed datasets without already knowing the exact query or source.
AI output needed context
A candidate or confidence score could accelerate review, but only when the interface clarified what the system detected and how to verify it.
Decisions required a trail
Reviewers needed to preserve provenance, compare candidates, and construct a defensible sequence from the evidence they accepted.
02
Research and synthesis
I mapped the decisions investigators made, not just the features they used.
Workflow analysis
Mapped how users moved from raw media through search, comparison, validation, and case-building.
Data synthesis
Grouped recurring needs, breakdowns, system states, and evidence relationships into design themes.
Rapid prototyping
Made abstract AI behavior tangible enough to evaluate with product, engineering, and stakeholders.
Iterative validation
Refined navigation, density, terminology, and review patterns as workflows became clearer.
Research constraint
When direct access is limited by sensitive evidence and high-stakes environments, every available signal has to work harder.
I paired stakeholder expertise, workflow evidence, product behavior, technical constraints, and iterative validation rather than treating any single source as the complete truth.
Three tensions shaped the solution
03
Workflow model
A clear path from the first search to a reviewable timeline of events.
- 01Locate
Search across evidence and narrow the field.
- 02Compare
Review candidate moments side by side.
- 03Validate
Inspect confidence, source, and context.
- 04Connect
Relate people, objects, places, and time.
- 05Construct
Build a reviewable sequence of events.
Investigation 24-0187Review in progress
▶
▶
▶
▶
▶
▶
Selected evidence
▶
AI similarity92%
- Source
- Camera 02
- Captured
- Oct 12 · 19:42
- Match basis
- Appearance + movement
Reconstructed to communicate interaction patterns and information hierarchy; not a reproduction of proprietary screens or case data.
04
Human oversight
Confidence is a prompt for review—not a substitute for judgment.
Evidence
Show the original source, time, and surrounding context.
AI candidate
Make the system’s proposed match visible and comparable.
Confidence
Communicate degree—not certainty—with the basis for the score.
Human review
Accept, inspect further, or reject while preserving the decision trail.
AcceptInspectReject
AI state and scope are visible.
Scores retain their evidence context.
Language avoids false certainty.
People can inspect and override.
05
Design system
Reusable patterns helped investigators scan dense information without losing context.
Layer information by decision value.
Summaries support scanning; progressive detail supports validation without forcing a context switch.
Candidate 0889%
Camera 11 · 20:14Never let color carry status alone.
Labels, icons, hierarchy, focus states, and contrast reinforce confidence and review state.
✓
ReviewedAccepted by J. Rivera
VerifiedTeach one review pattern, then reuse it.
Common card, drawer, filter, selection, and timeline behaviors reduced relearning across workflows.
Result cardDetail panelTimeline event
Incremental delivery
Partnering closely with product and engineering allowed patterns to be introduced cohesively within technical constraints—not as a disconnected redesign.
Design contribution
From powerful AI capabilities to clearer, more reviewable decisions.
Tool-centered entry points and fragmented evidence context.
Workflow-centered navigation and connected evidence relationships.
AI results presented as isolated outputs.
Confidence, provenance, and human review designed as one system.
Patterns that varied across complex product surfaces.
Reusable interaction models that supported cohesive implementation.
This case study emphasizes verified design contribution and qualitative product change; confidential business measures are intentionally omitted.
06
Reflection
People trust AI more when the interface is honest about uncertainty.
The biggest shift was treating explanation and human intervention as part of the workflow—not as disclosures added at the end.
That lesson applies beyond investigations: when AI influences an important decision, people need to understand what it found, why it surfaced the result, and when they should step in.