Faster quoting engine
1 hr → 3 min per job
Scheduling accuracy
Jobs finish in quoted window
Support deflection
Automated via SMS AI agent
Marketplace fill rate
vs ~70% industry
Fill rate and quote-accuracy metrics tracked after CV, OR, and SMS shipped; decision-ops patterns powered Command/Qortex review.
The Problem
Moovez, a last-mile delivery brokerage, was quoting jobs manually. Every estimate required a dispatcher to spend an average of one hour on phone calls, visual guesswork, and spreadsheet arithmetic.
The quotes were unreliable, leading to billing disputes and unpredictable scheduling. The support burden grew linearly with volume, creating a bottleneck that prevented the company from scaling efficiently.
The deeper failure was measurement: quote drift, noisy escalations, and dashboards that never wrote back to accuracy. Shipping faster quotes without decision-ops would have hidden margin harm. That gap is what ChurnOS patterns address: ranked attention, outcome capture, and YAML-governed verdicts instead of ad-hoc heuristics.
The Approach
Instead of patching the existing process, I rebuilt the estimation and operations workflow from first principles. By leveraging historical job data, computer vision, and operations research, I defined three interconnected systems.
The goal was to transform a manual, subjective process into an algorithm-grounded, automated engine that could be commercialized as a standalone SaaS: Quotely.
Shipping the three systems surfaced new failure modes: quote drift, noisy escalations, and success metrics that never wrote back to accuracy. I applied decision-ops patterns from ChurnOS to the production layer: a ranked attention queue, outcome capture with write-back, and YAML-governed verdicts instead of ad-hoc heuristics.
Three Interconnected Systems
Start: instrument vision with unknown-fallback leaderboard. Replaced manual inventory calls with a CV model. Customers photograph items; the model estimates volume and complexity to generate instant quotes.
Decide: verdicts grounded in job actuals. A predictive framework solving duration, optimal mover count, and 3D bin packing, trained on years of historical job performance data.
Decide + learn: missed vs noisy escalations. Built an automated SMS agent handling bookings, schedule changes, and queries, integrated directly with the operations backend.
Start: instrument vision, OR, and SMS. Decide: weekly ranked review and attention queue for what to fix first. Learn: eval harness and outcome write-back into quote accuracy.
Profitability closed the loop: MAE on quote vs actual, ±15% tolerance bands, and an outcome-capture SLA so billing disputes became measurable harm signals instead of anecdote.
Before → After decision ops
Production application at Moovez.
| Area | Before | After |
|---|---|---|
| Weekly review | Run-success dashboards | Ranked attention queue |
| Harm signal | Billing disputes, support volume | MAE, missed escalations, fill-rate drops |
| Policy | Ad-hoc heuristics | YAML SLAs, Command/Qortex playbooks |
| Retention loop | Ship and move on | Outcome flywheel → retrain/eval gate |
How I work
I define the metric first, then build until it moves. At Moovez that meant fill rate and dollar impact per quote, not feature count. I don't hand off work: I own it until it's measurably better.
I go solo from blank slate to working system, then design the handoff so the thing keeps running without me.
Speed matters at early stages. I ship, instrument, learn, and deprecate what doesn't work, with no attachment to the artifact. Decisions and outcomes have to write back to accuracy and escalation quality, not stop at dashboards.
Commercial outcomes
| Metric | Before | After |
|---|---|---|
| Time to generate a job estimate | ~1 hour | 3 minutes |
| Marketplace fill rate | ~70% industry | ~93% |
| Quote vs actual | Unmeasured | MAE + ±15%; outcome capture SLA |
| Jobs completing within quoted window | Unmeasured / High variance | 85% within range |
| Support tickets requiring human touch | ~100% of volume | ~50% deflected |
| Cost accuracy | Manual guess | Algorithm-grounded |
How accuracy closed the loop
Command/Qortex ops console, four-layer eval harness, and agent quality gates turned shipping velocity into measurable quote accuracy.
Weekly review surfaced what to fix first: fill-rate drops, quote drift, and missed escalations instead of run-success vanity metrics.
Regression, golden-set, shadow, and production sampling gates kept CV and OR changes from shipping blind.
SMS agent outcomes wrote back to training data. Missed vs noisy escalation labels fed the next eval cycle.
Growth systems beyond the product surface
I've also worked on the GTM side of product: using customer feedback, usage patterns, automation, and lead-routing workflows to improve retention, conversion, and ARPU. At BVXpress, the same decision-ops habit applied: define the metric, instrument, then automate the handoff.
Built personalized onboarding, usage-research surveys, marketing automation, and a new knowledge-base site for 1,400+ users.
Launched a 6-question Course Fit Calculator that scored prospects, wrote intent data to Zoho, and routed Tier-A leads to SDRs.
I map user journeys, identify system hand-offs, automate APIs and webhooks, then segment lifecycle messaging around measurable lift.
Other Project Work
End-to-end platform for microgrid siting and LCOE analysis. Generates automated Bill of Materials (BOM) for infrastructure projects.
Reduced false-positive alerts by 85% for a student-safety product deployed across 35 school districts.
Built the product function from scratch at an M&A firm, launching five products across research, workflow, and growth operations.
Useful at the messy stage.
I'm interested in roles where product judgment, engineering execution, and measurable outcomes all matter, especially agentic systems that need decision-grade ops.