The problem
Restaurants lose staff time and orders on the phone. Their owners need an ordering product that fits the business workflow, handles menu details, and responds quickly enough to keep a conversation moving.
Winning an early customer meant understanding that workflow, building a useful product, and supporting it after the pilot.
What I owned
I worked across every early stage: customer discovery, prototype, founder-led sales, pilots, deployment, and support. Pilot feedback from restaurant owners became small, scoped product changes.
My engineering work covered the voice stack, hybrid menu retrieval, model adaptation, evaluation, and infrastructure.
The engineering decisions
Combine retrieval signals. Menu lookup needs to handle exact names, noisy speech, and semantic intent. I built hybrid retrieval on pgvector with HNSW, metadata filtering, phonetic matching, and reranking.
Adapt models where useful. QLoRA and instruction tuning on SageMaker supported domain adaptation of open models including Llama and Gemma.
Engineer for conversation speed. State management, orchestration, and caching were part of the voice experience. The product needed to keep moving while preserving the order context.
Use a release gate. LLM-as-a-Judge evaluation and retrieval tests supported product iteration, with Kubernetes and Terraform for releases.
Discovery → pilot → observed friction → a scoped change → another customer conversation. Technical decisions were tied to whether the restaurant could depend on the product.
What I measured
I tracked business adoption alongside system behavior: pilot conversion, retention, and order completion. These answer different questions. A technically successful interaction does not, by itself, prove a product is worth paying for.
Retrieval tests and release evaluation supported the technical side of that feedback loop.
What changed
The startup grew to 12 paying locations, converting 80% of pilots to paid customers, with 100% retention during my tenure. Order completion reached 98% of qualified calls.
The team won the IDEA Lab Venture Competition. This work is separate from Checkmate’s VoiceBite pilot; the metrics describe different products and deployments.
Business and product outcomes refer to the April–October 2025 founder period. They are historical results, not a claim about the startup’s current performance.
A public engineering summary. Results are scoped to the project and period described; implementation details are summarized at a high level.