Engineer.
Operator. Translator.
I get sent in when the problem is not yet a spec. I sit with the people doing the work, map what is actually happening, and build agentic workflows that run in production.
Half of the week is discovery inside a client's systems. The other half is building the agent logic, the integrations, and the evaluation that proves it works. I do not stop at the prototype.
Shipped, live, and running right now.
Systems I designed and built that real people use every day. Solo or as technical lead, from empty repo to production.
StratAIgy Voice Assistant
AI phone answering platform for service businesses. Auth, onboarding, voice agent, billing, and analytics, shipped end to end.
- Webhook-driven state: call events and subscription events sync into Postgres, driving real-time access control
- Voice agent layer: trained per business on services, hours, pricing, and FAQs; books appointments, captures leads, escalates to a human
- Integrations shipped: Follow Up Boss, BoldTrail, Calendly, Google Calendar OAuth
- Guided 5-step onboarding with sub-10-minute setup, tiered plans with trial and upgrade paths
ClaimRx
Denied insurance claim recovery for small medical practices. Ingests claim data and generates payer-specific appeal letters with medical necessity arguments.
- Structured LLM generation: streaming output with CPT and ICD-10 citation plus payer-specific language, running at roughly one cent per letter
- 14 features shipped: JWT auth with auto-logout, claims CRUD and filtering, appeals lifecycle state tracking, analytics with drill-down
- Regulated data from day one: HIPAA compliance documentation and vendor BAA pipeline completed before pilot
- Five-table Postgres schema, Node and Express API, GitHub CI/CD auto-deploy
Aurum
Multi-tenant AI voice concierge platform with lead capture and per-client dashboards. Empty repo to a live external client in three days.
- Tenant isolation enforced at the database: 22 row-level security policies across 11 tables plus a platform-admin role architecture, verified against a real external account
- End-to-end pipeline: kiosk to API route to service-role write to Postgres to tenant-filtered dashboard
- Voice stack: Claude conversation layer, ElevenLabs speech with a pronunciation normalization pipeline, brand-safety constraints on agent output
- Documented SQL provisioning pattern so new tenants onboard repeatably
React Migration and Cutover
WordPress to React cutover on legacy shared hosting with a cascade of production blockers and an urgent deadline. Diagnosed and shipped in one evening.
- Diagnosed single-page-app routing failures and shipped an Apache rewrite config without touching application code
- Traced asset load failures through the browser console to a build-output versus server-structure mismatch
- Stripped 150+ lines of injected third-party markup and CORS-erroring analytics from production HTML
- Zero console errors at handoff
Contract and Billing Pipeline
Contractor running estimates, scope agreements, contracts, and HOA releases entirely by hand with no pipeline visibility. Zero to production in under two weeks.
- Stage-triggered document engine: a CRM stage change fires a webhook into a router that dispatches the correct API call per document type
- Bidirectional sync: an envelope-completion scenario polls signing status and writes back to the CRM, so no human updates state manually
- Accounting integration: QuickBooks connected to the sales pipeline, eliminating duplicate entry
- Four document types automated, three systems integrated, zero manual sends
Genesis Qualification Agent
Site-wide agent handling lead qualification, booking, referral triage across seven partner categories, and handoff to a human.
- Intent routing: classifies visitor intent, detects referral category, books directly, or hands off to a human when confidence is low
- Retrieval layer: knowledge base built by site crawl with scheduled weekly re-indexing to prevent answer drift
- Validated against five scenarios: qualified lead, unqualified visitor, referral, human handoff, partner application, all passing
Client Operations Agents
Three agents running on schedules against real client systems. Full breakdown in the Agents section below.
- Daily brief: reads tomorrow's calendar, then fans out per client across meeting transcripts, email, CRM, and internal chat into a structured pre-call brief
- Follow-up: pulls completed transcripts, cross-checks unanswered threads, drafts recap emails. Drafts rather than sends, as a deliberate human checkpoint
- CRM writeback: logs structured notes and lifecycle fields, diffing against current values so it writes only on actual change and re-runs stay safe
Prospecting and Sequencing Pipeline
End-to-end outbound system from profile definition through sequencing, follow-up, and logging. No manual touchpoints once launched.
- Targeted prospect sourcing with verified contact enrichment, filtered against a defined customer profile
- Multi-touch sequence with branching follow-up logic and automatic suppression on reply
- Every touch and outcome written back to the tracking layer automatically
How I run an engagement.
The same loop whether the client is a four-person contractor, a seventeen-agent brokerage, or an enterprise team with four stakeholders who want different things.
Embed
Get into their actual systems. Not a discovery deck. I ask for access and go find what is broken myself, then come back with specifics.
Map
Document the process that exists, including the spreadsheet nobody mentions and the step someone does from memory every Tuesday.
Split
Decide what the agent automates, where a human stays in the loop, and what happens when the agent is wrong. The highest-judgment call in the build.
Build
Tool calling, routing, retries, failure recovery, logging. Then the evaluation layer that earns the right to run on real volume.
Hand over
Documentation written so nothing is assumed. The client team operates and extends it after I am gone. Patterns come back as reusable pieces.
Embedded in 50+ client organizations.
Speculo builds an AI voice agent for real estate teams. My half of the job is forward deployed: go into each client's stack, find why the AI is not producing outcomes, redesign the workflow around it, build the integration, and train the humans who work alongside it. Nine CRM platforms, each with its own constraints and failure modes.
The client who had already quit
Months of poor results under the previous rep. She had stopped responding and mentally written the product off.
- Asked for direct CRM access instead of running a status call. Went in, found what was actually broken, came back with a specific plan
- Traced an integration gap between two of her systems that was silently dropping lead activity signals, leaving the agent blind on a whole class of lead, then built the workaround
- Built delay logic on new leads, handoff visibility lists, and opt-in flow configuration
- Raised the meeting cadence unprompted, because monthly would not surface problems fast enough on an account in that state
27,000 dormant records
A team brand new to their CRM sitting on a massive dormant database with no automation infrastructure.
- Diagnosed reported product bugs as login and routing configuration errors, which changed the entire remediation plan
- Designed a dual revive and convert strategy across 27,000 contacts at 250 to 300 calls per day over 60 days
- Built buyer and seller segmentation, mass tag architecture, and the full automation layer from zero
Four stakeholders, one bake-off
Enterprise client running our agent head to head against a direct competitor, with an inside sales team openly preferring the other product.
- Ran consecutive sessions across the owner, the inside sales lead, the technical admin, and internal product
- Refused to argue the comparison and committed to instrumenting it instead, with lead volumes and outcome rates for a fair evaluation
- Tuned agent call timing and exclusion criteria directly from operator feedback
- Closed a product visibility gap surfaced mid-meeting
Migration and cutover work
The unglamorous reality of integrating across platforms that were never designed to coexist.
- Sequenced an API activation by disabling the existing automation triggers first, preventing duplicate-fire conflicts during cutover
- Designed a five-trigger bridge-tag strategy for a constrained CRM, and validated it as buildable with the integrations team before presenting it to the client
- Ran a competitor-to-Speculo text transition on a live account without dropping in-flight conversations
- Stage architecture that excludes post-appointment stages, so the agent never calls an active client
Turning a broken human process into an agentic workflow.
Everyone assumed it was a UI problem.
Our AI was producing qualified handoffs. Humans were not acting on them. Product asked me to redesign the handoff experience. The real problem was not the interface. Nobody had decided where the agent stops and the human starts.
DISCOVERY
Read the full product playbook and every line of beta feedback. Mapped the real failure points: too many steps, mobile-first users on a desktop-designed flow, and a whole class of handoff people were completely blind to unless they logged into a dashboard they never opened.
THE AUTOMATE / HUMAN SPLIT
The agent owns qualification, timing, routing, and state. The human owns the claim and the relationship. A wrong claim is expensive and a bad first human contact is not recoverable, so that line is where it is on purpose.
INTERACTION DESIGN
Two-part flow. A structured alert with exactly one primary action, then the agent messaging conversationally, where the person's natural-language reply is the claim. No app, no login, no form.
ROUTING ARCHITECTURE
Four-layer distribution stack: inside sales routing, then source-group filter, then specialization opt-in, then first-to-claim or round-robin. Identical interface for a solo operator or a team of hundreds.
STATE AND DATA MODEL
Eight explicit states with matching CRM tags covering claimed, appointment set, appointment met, closed, expired, and released, so every handoff has an observable terminal state. A handoff that quietly dies is worse than one that fails loudly.
CONVERSATIONAL RETRIEVAL
Specced two-way question and answer against full conversation history, CRM activity, and lead behavior, so a person can interrogate the agent's reasoning on any lead inside the same thread.
Delivered: a 19-section interactive PRD with role filters for product and engineering, open questions flagged for decision, eight confirmed beta issues traced to root cause, explicit out-of-scope boundaries, and a build-readiness indicator. Co-presented to product and engineering. Designed forward-compatible with an emerging behavioral signal classification system, so the routing layer absorbs new signal types without a redesign.
Orchestration, evaluation, guardrails.
Multi-tool agents running on schedules against live business systems, plus the evaluation and safety layers that make them trustworthy enough to leave running unattended.
| AGENT | WHAT IT ORCHESTRATES | ENGINEERING DETAIL |
|---|---|---|
| Daily Brief | Reads tomorrow's calendar, then fans out per client across meeting transcripts, email, CRM records, and internal chat | Multi-source tool calling with per-entity fan-out, emitting a structured pre-call brief with open action items. Empty upstream has to mean empty, not failure. |
| Follow-Up | Pulls completed meeting transcripts, cross-checks for unanswered threads, drafts recap emails with action items for both sides | Runs twice daily. Writes drafts rather than sending. A deliberate human checkpoint, because the cost of an agent sending a wrong client email is asymmetric. |
| CRM Writeback | Logs structured meeting notes to contact records and updates lifecycle fields | Diffs against current values and writes only on actual change. Idempotent by design, so re-running never corrupts account history with noise. |
| Call Grading | Evaluates conversation quality against a rubric | Model-as-judge with an explicit anti-pattern guard: absent evidence is never scored as poor performance. Separates per-call signal from account-level trend to prevent false negatives. |
| Market Intel | Scheduled industry and competitive digest | Scoped retrieval with an explicit context firewall, so unrelated domains never bleed into output. |
GUARDRAILS I BUILD BY DEFAULT
Context firewalls between domains that must never mix. Explicit boundaries on what a system does not know, so an agent declines rather than fabricates. Human review on anything client-facing.
EVALUATION STANCE
Rubric-based grading with documented failure modes. The most common evaluation bug I design against is silence read as failure. Missing data is missing data, not a bad score.
SENSITIVE DATA
HIPAA-compliant workflow design with full compliance documentation on ClaimRx. Consent and outbound-contact constraints encoded directly into agent logic across the client book.
Systems I have connected in production.
Auth flows, webhooks, rate limits, pagination, duplicate-fire prevention, and upstreams that go down without warning.
AI AND VOICE
APPLICATION
DATA
ORCHESTRATION
CRM AND OPERATIONAL
DOCUMENT AND PAYMENT
INFRASTRUCTURE
PRACTICE
What the work produces.
Unprompted, and in order.
Rep one showed up prepared. Rep two took the reins. That's exactly what our clients need. Not "let's wait and see what happens."
She said here's my plan. I'm getting into your CRM myself. I'm going to see what's broken, and I'm reporting back to you. Then she put a weekly meeting on my calendar. Not because I asked.
Assigned lead on the company's most complex enterprise account. Tasked directly by product leadership to redesign the agent handoff system. Introduced AI tooling to the team, which drove company-wide adoption for client work and documentation.
My socks have been completely knocked off by the level of service that you provide. It's making a huge difference in my experience.