Work & Case Studies
A generative AI-powered intent capture agent — designed to understand callers the way they actually speak.
Executive Summary
Traditional, deterministic voice automation relies on rigid menus and exact keyword matches — leading to high failure rates when customers speak naturally. I led the business and technical feasibility analysis to evaluate migrating a legacy, rule-based Main Menu routing agent to Wayfinder, a Generative AI-powered intent capture agent.
By analyzing production routing data, I built an evidence-based business case proving that a semantic-based GenAI model would capture conversational, complex caller intents that the deterministic system completely missed — projecting a significant lift in containment and a reduction in manual agent transfer costs.
Showcase · User Persona
Traditional IVR fails the moment a caller speaks naturally. This persona defines the real user Wayfinder was built to serve — and makes concrete why the legacy system consistently let her down.
Hates listening to long, scripted menu options. Experiences cognitive overload when forced to remember options 1–5 while multitasking.
"My statement looks wonky" crashes regex patterns. The system loops her through error prompts rather than attempting to understand her intent.
When she finally reaches a live agent, zero context transfers. She re-states her full issue from scratch — adding frustration and handle time.
Designing around Sarah's actual behavior — not idealized behavior — was the foundation of the Wayfinder business case. Every design decision traces back to the gap between how she naturally speaks and what the legacy system could process. This persona brief aligned the entire product team before a single line of code was written.
Showcase · Customer Journey Map
The same four interaction stages — Initiation, Intent Capture, System Processing, Routing Outcome — play out in radically different ways depending on which engine is handling the call.
Mapping both paths side-by-side turned an abstract technical argument into a concrete business case. Stakeholders who saw the legacy path's emotion row — from Neutral to Frustrated Transfer — immediately understood the cost of inaction. The journey map became the centerpiece of the executive presentation.
Showcase · Sample Dialog
This annotated script spec demonstrates how Wayfinder handles ASR transcription errors — the real-world condition where speech recognition mishears words — and still extracts the correct intent from context.
This scenario proved the core technical thesis: Wayfinder's semantic model extracts correct intent even when the underlying ASR transcription is corrupted. A deterministic system requires perfect keyword matching — Wayfinder requires only coherent sentence structure. This single scenario was the decisive evidence in the architecture review board presentation.
Showcase · Call Flow Architecture
This architecture diagram shows exactly where the two systems diverge — and why that divergence produces fundamentally different outcomes for callers who don't speak in keywords.
Scanned for explicit keywords like "Pay Bill" or "Balance." The colloquial word "wonky" crashed the regex patterns.
Semantic embeddings analyzed the sentence as a whole. "statement online" + "amount looks wonky" = 94% contextual probability to billing.
Caught the word "driving" and mistakenly routed the customer to Roadside Assistance instead of tech support.
LLM reasoning isolated secondary noise ("driving right now") and extracted the true core intent: "heater clicking," "freezing."
Matched two conflicting intents simultaneously and forced a generic "I didn't quite catch that" prompt.
Identified the primary action item, routed to card activation and passed a Stateful Context Tag so the downstream agent handles the address update without asking the caller to repeat.
These three scenarios — drawn from production call logs — formed the quantitative backbone of the architecture review. Each one represents a failure mode the legacy system could not recover from. Wayfinder resolved all three through a single architectural change: replacing keyword extraction with full-sentence semantic evaluation.
Showcase · Voice Wireframe
A Voice Wireframe defines the system's exact utterances, the expected caller inputs and the technical processing logic — before engineering writes a single line of code.
This wireframe was the primary handoff document to the engineering team. It defined every decision branch, processing step and fallback behavior before development began — reducing scope ambiguity and cutting the first-sprint rework cycle by an estimated 40%. The annotated architecture also became the training document for the QA team during acceptance testing.
Showcase · Prompt Engineering Framework
A repeatable design methodology for engineering prompts that are reliable in production — built from the intersection of conversational design craft and generative AI architecture. Every principle maps directly to real failure modes observed in enterprise deployments.
Most prompt failures aren't model failures — they're design failures. The C.L.E.A.R. framework applies the same rigor used in enterprise conversation design to the challenge of shaping LLM behavior at scale.
Define the system's role, constraints and persona before any task instruction. A model that doesn't know who it is will hallucinate an identity — and that identity may not align with your brand, your compliance requirements or your users' expectations.
Design for how people actually speak — not how you wish they would. This means accounting for slang, incomplete sentences, ASR transcription errors, code-switching and multi-intent utterances. If your prompt only works with clean input, it won't survive contact with real users.
Every prompt must have a graceful fallback — no dead ends. When the model can't determine intent with sufficient confidence, it should ask a targeted clarifying question, not loop a generic error. When it can't resolve the issue, it should hand off with context — not abandon the user.
Set explicit confidence gates that determine when to route, when to clarify and when to escalate. Don't let the model guess at 51% confidence — define the floor below which human judgment is required. Thresholds are a design decision, not a model parameter.
Prompts are never final. Build in a structured testing, red-teaming and iteration cycle before launch — and a monitoring and optimization process after. Production data will surface failure modes that no test suite predicted. The refinement loop is what separates a deployed prompt from a production-ready one.
Why This Framework
Most prompt engineering frameworks focus on technical output quality — does the model return the right answer? The C.L.E.A.R. framework focuses on the human experience of interacting with a system shaped by that prompt. It asks: does the caller feel heard, understood and moved forward? That shift in perspective is what makes AI-powered conversations feel less like software and more like service.
The C.L.E.A.R. framework was developed from patterns observed across enterprise voice AI deployments. Each principle addresses a specific, documented failure mode. When applied to the Wayfinder architecture, the framework produced a system that correctly identified caller intent at 94% confidence on corrupted ASR input — a result that deterministic keyword matching couldn't approach.
Like what you see?
These showcases represent the craft and rigor I bring to every engagement. Ready to talk about your project?