THE PROBLEM
Enterprises racing to adopt AI. Employees pasting sensitive data into ChatGPT. Traditional DLP either blocking AI outright or missing natural-language prompts entirely.
When Unbound joined Y Combinator's Summer 2024 batch, the enterprise AI security category was barely a category. Enterprises were racing to adopt ChatGPT, Claude, and Copilot. Employees were pasting source code, customer data, and credentials into prompts every day. Traditional data-loss-prevention products either blocked AI outright, killing productivity, or failed entirely because they could not parse natural-language prompts the way they parsed structured database queries. Unbound's thesis was straightforward: enterprises will adopt AI no matter what security wants. The job is to make adoption safe by intercepting, analysing, and controlling every AI request at the right layer of the stack. Building UI for that thesis meant inventing patterns: there were no policy editors, risk-tier visualisations, or agent-access matrices to borrow from competitors.
The original product UI looked like a generic SaaS dashboard. Enterprise security buyers want density, precision, and trust.
The MVP Unbound had when Say Design joined was directionally working but visually undifferentiated. Flat layouts, generic component library, no system underneath. Enterprise security buyers are a specific audience. Chief Information Security Officers and security architects do not want playfulness. They want density, precision, and trust cues across every artifact. Risk scores, leak counts, audit trails, and policy histories all need to communicate enterprise credibility on first glance. A flat MVP dashboard reads as a side project to that audience. The first call after joining: rebuild the entire dashboard on top of a modern B2B design system from the ground up, before adding new features, before considering the next pivot, before any marketing site or pitch deck work.
AI visibility platform, then AI Gateway, then Agent Access Security Broker. The design system had to flex without rebuilding.
Across 20 months, Unbound pivoted three times as the company searched for product-market fit. Pivot 1 was the original AI visibility platform with a browser extension and enterprise dashboard intercepting employee prompts in ChatGPT, Claude, and Copilot. Pivot 2 moved up the stack to an AI Gateway intercepting Cursor, Roo, Cline, and internal document copilots, enforcing per-team policies and routing requests across internal, external, and open-source models. Pivot 3 reflected the industry shift from chat-based AI to agentic AI: an Agent Access Security Broker operating at the agent-access layer with run-through customisable policies enterprises apply to every AI agent their teams use. Each pivot tested whether the design system was a help or a cage. A well-architected system flexes. A brittle one forces rebuilds. The system survived all three.
THE DESIGN JOURNEY
Twenty months as founding designer, daily 11 PM IST stand-ups, sales-driven feature flow.
Embedded with Raj and Vignesh as part of the team, not as an external agency. Linear tickets alongside engineers. Customer signals turning into features in days, not quarters.
Embedded model with daily founder calls. Direct communication on every product and feature request. Linear as the work tracker.
Say Design joined as founding designer in June 2024 during Unbound's YC Summer 2024 batch and stayed embedded for 20 months. The operating model was deliberate: not an external agency relationship, but a working role inside the Unbound team. Daily stand-ups at 11 PM IST with Raj as CEO and Vignesh as CTO. Direct founder communication on every product and feature request, no PM layer between design and engineering. Linear carried the work tracker alongside the engineering team. Occasional weekend turnarounds when a pitch deck needed to land before a Monday investor call or when a launch customer demanded a fix on a tight timeline. Twenty months across three pivots without an account-management gap or a context-switching tax. The engagement was one of the longest-running founding-designer retainers in Say Design's portfolio.
Product requests came straight from Raj and Vignesh, who were collecting signals from sales conversations and customer feedback.
Most YC-stage startups run feature requests through a PM filter. Unbound did not. Sales conversations and customer demos surfaced requirements directly, Raj and Vignesh translated those signals into Linear tickets, and design picked up the work the next day with founder context already loaded. This worked because Raj's prior tenure at Palo Alto Networks and Imperva meant he could distinguish real customer pain from a tire-kicker request, and Vignesh's engineering rigour meant feasibility constraints arrived alongside the design brief. A typical week: a security architect at a Fortune 500 customer asks for a per-team policy override during a demo, the founders ship the requirement that evening, design wireframes ship in 24 hours, high-fidelity in 48, engineering picks up handoff on day three.
Three pivots tested whether the design system was a help or a cage. The system flexed every time.
A well-architected design system flexes through pivots. A brittle one forces rebuilds. Unbound's three pivots over 20 months were the live test. Pivot 1 to Pivot 2 changed the layer from browser extension to infrastructure gateway, but the underlying primitives held: dashboard layout, policy-editor patterns, analytics conventions, audit-trail components. Pivot 2 to Pivot 3 changed the mental model from prompt-level interception to agent-access mediation, but the same primitives extended: policies became agent-policies, dashboards added agent-specific dimensions, audit trails added agent-identity columns. The system was designed from day one to scale through Series B without a rebuild. The primitives that shipped in the original AI visibility dashboard are still shipping in the AASB.
THE SOLUTION
A modern B2B design system, a marketing site in Framer, a fundraising-grade pitch deck, and a multi-surface visual language tied together.
Product, website, pitch deck, and LinkedIn collateral all shipping under one designer with one visual language across 20 months.
Inherited a flat, generic MVP. Built a system the company could ship on through three pivots and a $4M raise.
The first major delivery was a complete rebuild of the dashboard on top of a modern B2B design system architected from scratch. Component primitives covering tables, charts, policy editors, audit trails, run-history surfaces, and agent-detail views. Token system for spacing, type, colour, and elevation defined semantically so the same colour token meant "high-risk" everywhere it appeared, not just in the original component it shipped on. Variant coverage exposed up front: every component shipped with its full state matrix including default, hovered, focused, pressed, disabled, loading, error, and empty. The dashboard rebuild became the foundation every subsequent feature inherited from. New analytics dashboards in Pivot 2 reused chart primitives. Agent-access matrices in Pivot 3 reused table and policy-editor primitives. The system absorbed three pivots without a structural rebuild.
Framer marketing site, fundraising-grade pitch deck for the $4M seed, LinkedIn cadence, brochures, event collateral.
The engagement covered every surface a YC-stage enterprise security company ships, all under one designer with one visual language. The marketing site at getunbound.ai was designed and built in Framer, multi-section, aligned to the enterprise AI security category positioning. The pitch deck used in the May 2025 seed raise was designed end-to-end: narrative arc structured for Race Capital and Wayfinder Ventures, single-idea slides, data visualisation at investor-grade precision, category positioning that staked Unbound's claim as the safety, observability, and cost-discipline layer for enterprise AI. LinkedIn content cadence shipped continuously, brochures and one-pagers for sales conversations, campaign creative for category-defining moments, and event collateral for security conferences where Raj and Vignesh were in front of CISOs. Multi-surface ownership in one engagement meant the visual language stayed unified across the moments customers actually encountered Unbound.
Writing a rule that says "block all prompts containing API keys except from Engineering using Cursor" is a real UX problem.
Two surfaces carried the most product-design weight across the engagement. The policy editor: writing a rule like "block all prompts containing API keys except from the Engineering team using Cursor" needed to be readable, auditable, and modifiable by non-engineers. The editor surfaced policies as plain-language rule statements with structured slots for conditions, actions, exceptions, and override paths, with a code-view available for power users. The analytics dashboard: risk scores, leak counts, model-routing stats, cost savings, and per-team breakdowns all competing for the same screen. Hierarchy decisions here are the product. Risk tier surfaces first as the headline metric. Cost savings surface second as the executive proof point. Per-team and per-agent drill-downs sit one click below. The dashboard reads as a security operations centre console.
THE KEY MOMENT
A design system built on day one to ship at $50M Series B scale.
The signature decision came in the first month of the engagement. The flat, generic MVP needed a rebuild, but the question was how deep to architect the new system. The shallow path was to ship a polished dashboard skin and move on. The deeper path was to architect a system on day one that could absorb whatever Unbound became. The deeper path won, and that decision compounded across 20 months. Component primitives were named by function not feature: a "policy editor" was named for the abstract job, not for the AI-visibility-pivot context it first shipped in, which meant the same primitive carried directly into the AI Gateway pivot and then into the AASB pivot without renaming or refactoring. Tokens were semantic, not literal: a colour was tagged "risk-high" rather than "red-500", which meant the same token applied across analytics dashboards, policy editors, and audit trails consistently as the surface area grew. Variant coverage was exposed up front: every component shipped with its full state matrix on day one, so engineering never had to ask design what the disabled-and-loading state of a button looked like. The system was designed for a company that did not yet exist, the version of Unbound that would close a Series B and ship to thousands of enterprise customers, not the version that was raising a YC pre-seed. Three pivots and a $4M oversubscribed seed later, the same primitives are still in production.
The work, in detail.
Dashboard surfaces across three pivots, the design system primitives, policy editor variants, analytics dashboards, marketing site sections, and the deck used in the $4M seed raise.
Unbound Security
industry:
Security, B2B SaaS, AI
SCOPE:
Website
,
Product
,
Creatives
duration:
2 Years
key outcome:
$4M
in funding
