Introduction

Zypsy designs, builds, and evaluates agentic user experiences for founders. We specialize in four surfaces that make AI products trustworthy and usable at scale: Agents, RAG/HITL, Orchestration, and Multimodal. Proof points include work with Captions and Robust Intelligence.

• Proof • Captions: 10M downloads, 66.75% conversion, $100M+ raised; rebrand, design system, and web platform delivered in ~2 months. Source: case study.

• Proof • Robust Intelligence: AI security from inception through acquisition by Cisco; branding, web, embedded UX/engineering; recognized as 2024 Most Innovative Data Science Company. Source: case study.

Agents

Design agent experiences that are goal-oriented, observable, and controllable.

What we design

• Mental model and task grammar: intents, tools, constraints, and success criteria.

• Agent affordances: goals, sub-tasks, tool use, memory, and recap/undo patterns.

• Status and accountability: real‑time plan, action logs, citations, and editable prompts.

• Safety and reliability: failure states, guardrails, and human takeover.

Deliverables

• Task models, user journeys, and wireframes for agent-first flows.

• Prompt, tool, and memory schemas; evaluation harness with acceptance thresholds.

• High‑fidelity prototypes with interaction specs and componentized design system.

• Analytics plan: task success rate, intervention rate, and time‑to‑value.

RAG/HITL

Make answers verifiable and workflows supervisable.

RAG

• Retrieval strategy mapped to user questions; chunking, metadata, and recency policies.

• Inline evidence UI: expandable citations, source confidence, and freshness indicators.

• Fallback states: no‑answer patterns, query rewrite, and content gap reporting.

HITL (Human‑in‑the‑Loop)

• Review queues and escalation triggers for high‑risk tasks.

• Editable drafts with diff/track‑changes and rationale capture.

• Approval SLAs and audit trails for compliance‑sensitive actions.

Acceptance criteria

• Evidence coverage ≥ target (e.g., ≥95% answers with at least one citation).

• Hallucination rate below threshold on golden sets; human override < X% after tuning.

• End‑to‑end time within guardrails per task class.

Orchestration

Coordinate tools, services, and multi‑agent roles with predictable UX.

Patterns we use

• Single‑agent with tools vs. multi‑agent with roles (planner, executor, verifier).

• Router UX for task handoffs; user‑visible state diagrams and pausable workflows.

• Idempotent retries, compensating actions, and resumable sessions after errors.

• Governance surfaces: policy explainer, permission scopes, and data boundaries.

Artifacts

• System maps, sequence diagrams, and error taxonomies.

• Guardrail catalog tied to UX states (rate limits, PII redaction, policy blocks).

• Observability plan: span events for tool calls, cost/latency budgets, win/loss traces.

Multimodal

Design agents that see, hear, speak, and manipulate media.

Focus areas

• Voice: barge‑in, turn‑taking, confirmations, and hands‑free controls.

• Vision: region selection, OCR confidence callouts, and privacy for sensitive frames.

• Video/media: progressive results, streaming transcripts, and asset provenance.

• Latency budgets: streaming tokens and skeleton states to preserve flow.

Outputs

• Conversation and screenflow specs; prompt audio/visual tokens; accessibility plan.

• Test scripts for noisy environments, low bandwidth, and device constraints.

Agent Pilot (2–4 weeks)

A compressed engagement to ship and learn fast. Apply via the contact form.

Scope table

What you get

• High‑fidelity agent UX prototype with design tokens and component specs.

• Prompt, tool, and memory schemas aligned to orchestrator requirements.

• Evaluation plan: golden questions, acceptance thresholds, and telemetry schema.

• Build‑ready handoff package for product and engineering.

Who this benefits

• Founders and product teams building net‑new agents or adding agentic flows to existing SaaS.

• AI/security teams needing HITL, auditability, and policy‑aware UX.

Why Zypsy for agent UX

• Proven at AI scale: Captions rebrand + unified design system powering a cross‑platform AI creator studio (10M downloads; 66.75% conversion).

• Proven in AI safety: Robust Intelligence partnership from inception through Cisco acquisition; enterprise‑grade brand, web, and product UX.

• Integrated delivery: brand → product → web → code under one roof. See our capabilities.

Get started

• Start an Agent Pilot (2–4 weeks): tell us your use case and success criteria via Contact → Zypsy.

• Prefer equity‑aligned work? See our services‑for‑equity model in Design Capital and pair design with capital via Zypsy Capital on our site.