Convin

The short version

Problem. QA logic was custom Python written per client, agents only heard what they did wrong after the call ended, and the LMS had under 30% completion.

Action. Designed all three core systems: a no-code Rule Engine, Live Agent Assist for mid-call nudges, and a Coaching & LMS rebuilt concept-first, while building and leading a five-person design team.

Result. Audit coverage grew from 2% to 85%, nudge latency fell from 4 to 8 seconds to under 1, LMS completion rose from under 30% to over 70%, and two enterprise accounts landed.

The impact

• 2% → 85% audit coverage
• 2 enterprise accounts landed
• 30% faster design-to-dev cycle

How the systems connect

Audit. Rule Engine

No-code logic builder that replaced custom Python written per client.

• 2% → 85% audit coverage
• 2 enterprise accounts landed

Assist. Live Agent Assist

Real-time nudges during live calls, powered by the same Rule Engine logic.

• 4–8s → <1s nudge latency
• ~2x faster agent reaction

Improve. Coaching & LMS

Turns audited calls into structured agent coaching and learning plans.

• <30% → 70%+ LMS completion
• ~4.5 days faster agent ramp

Audited calls feed structured coaching, closing the loop instead of leaving QA data unused.

Background

Convin builds voice AI for contact centers: call auditing, real-time agent assist, and post-call coaching. Joined as a Design Intern, promoted three times over 3.5 years into Senior Product Designer, eventually owning all three of the platform's core systems and the design team building them.

System 1: Rule Engine, audit coverage 2% to 85%

The Rule Engine's no-code logic builder on a laptop screen.

The problem. An enterprise client needed deeply conditional QA logic. The existing system was custom Python written per client: 7 to 14 days of engineering time per template, nothing reusable across clients. The real problem wasn't one client's request, it was that Convin had no scalable way to configure complex QA logic without engineering being the bottleneck every time.

What was considered

Salesforce's own workflow engine was flagged directly in research as powerful but painful to use, the thing specifically to avoid. The first drag-and-drop version broke because most real users ran Windows browsers at 125 to 150 percent zoom, making the canvas too tight for reliable drag actions.

What was let go of. Product wanted to launch faster with fewer features once the build ran past 120 days, well past the original MVP scope. Pushed back: shipping it raw would just move the bottleneck, not remove it. Launch was delayed a few weeks to build reusable blocks and proper test and debug views; some complex features were cut outright rather than shipped half-working.

Leading it without the title. Sole designer on Rule Engine, run like a design lead: mentored a visual designer on systematic thinking, guided an intern who now owns Rule Engine iterations full-time, ran internal walkthroughs and live client training calls.

Result. Audit coverage expanded from 2% to 85%. Helped land 2 major enterprise accounts. Cut resolution times by 30%.

System 2: Live Agent Assist, from reporting on agents to shaping them mid-call

The Live Agent Assist panel showing real-time nudges on a desktop.

The problem. Convin's product told agents what they'd done wrong after the call ended. Useful for coaching, useless in the moment that actually mattered, while the agent was still on the phone with the customer.

The reframe. This wasn't a feature request, it was a category question. Post-call analysis is a compliance and reporting spend; real-time assist is a revenue-enablement spend, a different budget line and a different buyer entirely. Repositioning Live Agent Assist as real-time enablement rather than faster QA changed who Convin was actually competing for, and opened a meaningfully larger opportunity than post-call QA alone.

The latency problem. Early builds surfaced nudges 4 to 8 seconds after the relevant moment in a call, by the time a suggestion appeared, the conversation had already moved on. Not a polish problem, the difference between the feature working and not working at all. Getting this under a second took real collaboration with engineering, a new recording and processing approach to cut the delay. The design side of that work was deciding where in the interface a nudge needed to appear, how much information an agent could absorb mid-conversation without breaking their flow, and testing what "fast enough to be useful" actually meant on a live call versus a lab demo.

My role. Partnered directly with engineering on the latency fix, translating "will this feel instant to an agent on a call" into concrete interface and timing requirements. Owned the interaction design for how and when a nudge surfaces without disrupting the agent. Connected the system to the Rule Engine, so the same logic layer powering audits could power real-time nudges, one platform, two applications, rather than a parallel system built from scratch.

Result. Latency cut from 4 to 8 seconds to under 1 second. Roughly 2x faster agent reaction time during live calls.

System 3: Coaching & LMS, closing the loop from audit to actual improvement

The Coaching & LMS experience on a laptop.

The problem. Rule Engine caught what agents were doing wrong. Live Agent Assist nudged them in the moment. Neither helped an agent get better over time, that was still manual and inconsistent, dependent on how much bandwidth a manager had that week. Convin's early LMS attempt had under 30 percent completion, and coaching that didn't map back to what calls were actually revealing agents needed help with.

The insight. Research found the LMS was failing for a structural reason, not a content reason: it taught concepts after showing agents their mistakes, when the more effective order was concept first, mistake second. One participant's own words during testing confirmed it directly, seeing the correct approach before their own error made it land, rather than feeling like criticism. That sequencing change became the backbone of the redesign.

What didn't work. Two early bets didn't survive contact with real agents. Mock call placement, inserting practice calls too early in the flow, interrupted learning momentum rather than reinforcing it. TTS voice for training scenarios undermined how seriously agents took the content, replaced with real recorded call snippets. Both were cut rather than shipped half-working.

Result. Completion moved from under 30 percent to over 70 percent. Coaching time dropped roughly 40 percent as managers spent less time repeating fundamentals. New agents ramped roughly 4.5 days faster. A measurable CSAT uplift followed for coached agents, exact figures held under NDA.

Beyond the systems: building the team

Built and led a 5-person design team (2 Product Designers, 2 Visual Designers, 1 Intern) through Series A growth. Owned hiring end to end and built a full design ops process, including mentorship and learning plans for reports. Introduced design review, documentation, and QA standards that cut design-to-dev iteration cycles by 30%. Used AI tools to rapidly build Convin's initial component library, then partnered with a dedicated designer to refine it into a full design system.

At a glance

Role

Senior Product Designer (promoted 3x, Design Intern to Senior Product Designer, 3.5 years)

Team built and led

5 designers (2 Product Designers, 2 Visual Designers, 1 Intern)

Systems designed

Rule Engine, Live Agent Assist, Coaching & LMS

Audit coverage

2% to 85%

Enterprise accounts landed

2 major accounts, directly attributable to Rule Engine

CSAT / Retention

+27% CSAT, +25% retention across core AI systems

Design-to-dev cycle

cut 30% via review, documentation, and QA standards