Applied AI systems built around real workflows, deterministic controls, tool calls, human approval, evaluation, and explicit failure recovery.
recreatedA bounded AI pilot keeps allowed actions, evidence, and escalation paths explicit.
At a glance
What I owned
Period
January 2026 to present
Role
Co-founder; applied AI engineering lead
Team
Acelera team working with client stakeholders
Users
Operations and business teams evaluating AI-assisted workflows
Status
Active AI implementation studio; Clara Voice is a tested pilot prototype, not a production deployment
Contribution
Mapped operational workflows and built TypeScript and Python LLM orchestration with tool calls, deterministic rules, human approval, evaluation, and failure recovery.
Tradeoff
Let the model interpret natural language while deterministic software controls availability, required evidence, consent, allowed actions, and escalation.
Evidence
Clara Voice is covered by 95 automated tests. Client work remains NDA-safe and is described as discovery, prototype, and pilot handoff rather than production deployment.
System
Clara Voice
Control
Tools + deterministic rules
Review
Human escalation
Tests
95 automated tests
Sanitized engagement model
Scoping one observable constraint
Stakeholder interviews and current handoffs become a bounded AI pilot. The model interprets natural language, tools expose allowed capabilities, deterministic rules control decisions, and failures escalate to human review.
Workflow signals / sanitized
Repeated status entry across tools
Exception ownership not visible at handoff
Pattern condensed from stakeholder interviews
Current handoff
01Request arrivesInbox
02Context rebuiltObserved delay
03Owner assignedSpreadsheet
04Decision sharedChat
Pilot boundaryPreserve context at handoff 02
Owner
Operations lead
Signal
Time to assign
First release
One workflow
Client details removed. This is the engagement model, not a client artifact.
01
The brief behind the brief
Several engagements began with a request to use AI. I interviewed the people doing the work, mapped the handoffs and failure cost, and reduced the request to one workflow with an observable success signal.
recreatedDiscovery narrows a vague technology request into one observable constraint.
02
Clara Voice: language is probabilistic, actions are not
Clara Voice interprets natural-language appointment requests. Tool calls expose availability and workflow actions, while deterministic rules control required documents, consent, allowed transitions, and when the system must stop.
03
Failure handling belongs in the product
Provider failure, missing documents, unavailable slots, and clinical-risk signals route to explicit fallbacks or human review. The prototype is covered by 95 automated tests across orchestration, validation, and failure paths.
recreatedA bounded pilot makes allowed actions and escalation paths reviewable.
04
A public product, not a client claim
Brand Kit Generator turns an 8-stage interview into a decision tool. Each stage has a gate for rejecting vague positioning, voice, or visual choices before producing reusable brand outputs. The methodology and source are public.
TypeScript and Python orchestration connect LLM APIs, tool calls, voice providers such as ElevenLabs and Twilio, deterministic policy checks, evaluation fixtures, and human escalation paths.
06
Claim boundary
The healthcare work progressed from stakeholder interviews through a tested prototype and defined pilot handoff. This case does not claim a production deployment.
