Maya handles 70–80% of inbound support at first touch on one US utility. That’s 100+ emails per manager per day that never reach a human.

What it is

Two surfaces. A customer-facing assistant that handles inbound across voice, chat, SMS, email, and mobile, 26 languages. An internal console for human agents, plugged into Oracle CCS, that loads the context before they pick up.

Most support queries are repetitive. Bill questions. Outage status. Account changes. Maya handles the first touch. When it can’t, it hands off to a human with full context already loaded.

What I built

I was the only AI person at ESC Partners when this started. No handoff from a previous team. No prior codebase to extend.

Built it in 2.5 months. FastAPI backend, React frontend, AWS Bedrock for the LLM layer, DynamoDB for session state, n8n for orchestration, deployed on Render. Oracle CCS was the hardest part, that documentation isn’t designed for this kind of use.

The moment that made it real: watching an agent use the console for the first time with a live customer. They didn’t have to dig through three tabs to find the account. It was just there.

The number I can defend

70–80% first-touch coverage on one US utility. I can explain exactly how we measure it, what counts as handled, and where the edge cases live.

I will not give you a bigger number than that.

Stack

FastAPI, React, AWS Bedrock (Claude), DynamoDB, Oracle CCS, n8n, Render. Voice via third-party provider. SMS and email through dedicated channels. Mobile via customer widget.

Care coordinators were drowning in leads. Every inquiry got the same priority, a daughter researching for next year sat in the same queue as a son who needed a bed by Friday. No scoring. No triage. Just a spreadsheet and good intentions.

The problem

Assisted Living Locators connects families with senior care options. Their coordinators talk to hundreds of families a month. The issue wasn’t volume, it was signal. They couldn’t tell which leads needed a call today and which could wait a week. That meant slow responses to urgent families and wasted time on cold ones.

What I built

Two surfaces. A scoring flow that ranks and matches families to care options based on urgency, geography, care level, and budget. And a console the coordinators actually use when they pick up the phone, context loaded, match ranked, notes from the previous call already there.

I owned the data layer behind both. The scoring model pulls from intake forms, follow-up signals, and care-facility availability. The console reads from the same source so coordinators never have to cross-reference tabs.

How I worked

Remote, three-month engagement. I was the builder, not the advisor. Shipped production code, not a strategy deck. Stood up the data layer, wired the scoring logic, and built the console interface. When I left, both surfaces were live and in daily use.

What I’d do differently

The scoring weights were set by gut and coordinator feedback. With more time, I’d have built a feedback loop, track which scored-high leads actually converted and retrain the weights against real outcomes.

Zero product. Zero customers. Zero codebase. I was the first and only PM. My job was to figure out what to build and make sure it shipped.

The problem

Employee coaching is expensive. A good facilitator costs real money per session and can only run one at a time. Most companies skip it entirely or run it so rarely it doesn’t stick. Vollie’s bet: a voice AI that can moderate a coaching session, keep it on track, and produce the summary a manager actually reads.

What I built

A voice AI that runs the session live, asks the questions, manages turn-taking, keeps the conversation on topic, and handles awkward silences without making them worse. After the session ends, it writes the report: sentiment analysis, engagement scores, commitments made, and follow-up items.

I scoped the first three features. Cut everything that didn’t fit a first release. Wrote the eval suite that measures whether the AI’s moderation actually improves conversation quality or just interrupts it.

How it went

Zero to first client in seven months. The hardest part wasn’t the AI, it was defining what “good moderation” means well enough to evaluate it. A facilitator who talks too much is bad. One who says nothing is worse. The eval suite had to capture that spectrum.

What I learned

Founding PM means you’re the person who says no. The team had ideas for real-time translation, async coaching, Slack integration, manager dashboards. All good ideas. None of them mattered until the core session worked. I cut the roadmap to three features and didn’t add a fourth until the first three shipped.

Two products. One year. Built both while finishing a master’s at Simon Business School.

VentureForge

Strategy simulations, graded by AI. Students get a scenario, a market entry, a pricing decision, a pivot call, and work through it. The AI evaluates their reasoning, not just the answer. Did they identify the right tradeoffs? Did they consider second-order effects? The grade comes with an explanation of what they missed.

I built this for the Ain Center for Entrepreneurship at Simon. The goal was practice reps. MBA students read plenty of case studies. They rarely get to make the call and see it graded in real time.

Validation tool

The second product. Founders come in with an idea. The tool scores it on market fit, competitive landscape, and execution feasibility. Then it generates personas, synthetic target users based on the idea’s market, and runs simulated user interviews with them.

No researcher needed. No two-week wait for five interviews. A founder can test an assumption in minutes and get a signal before they spend a month building the wrong thing.

How I worked

Product lead on both, simultaneously. Scoped features, wrote specs, shipped. The hardest constraint was time, I was a full-time student building production tools. That forced a discipline I still use: if a feature can’t be scoped to two weeks, it’s too big. Break it down or cut it.

What matters about this

These weren’t class projects. They’re tools the center still uses. Students and founders interact with them. That’s a different bar than a grade, the code has to work when you’re not watching.

Two companies. Five years. Two countries. I learned product by doing everything around it first.

CloudApproach

Started in London, expanded to Dubai. AI tools for real estate marketing, automated listing descriptions, market analysis, and campaign optimization. The bet was that real estate agents spend too much time writing copy and not enough time selling. We automated the copy.

I built the team. Eight people at peak. Closed deals with agencies across the UK and MENA. Hit the product-market fit, scaled it, and sold the company. The combined acquisition value across both companies was $575K.

Approachables

The second company, run in parallel. A product-led marketing agency, we didn’t just consult, we built the tools our clients used. Positioning, landing pages, conversion flows. Revenue influenced across the full run: roughly $5M.

Sold this one too.

What I actually learned

Hiring is the hardest product decision nobody calls a product decision. Every person you bring on reshapes what you can ship. Get it wrong and you lose months. Get it right and features appear faster than you planned them.

Payroll. Pricing. Legal. Cash flow. The parts nobody puts in a case study because they’re not glamorous. But they’re the reason most startups die. I ran both companies through all of it.

Why it matters now

I don’t call myself a founder anymore. I call myself a builder. But the founder years are why I think about products differently. I’ve signed the checks. I’ve made payroll when the bank account was thin. When I say I understand the business side, I mean I’ve lived it, not read about it.

The City of Rochester had a fraud model at 62% accuracy that nobody trusted. When it flagged a case, the loan officers couldn’t tell why, so they second-guessed every decision.

What I built

An explainable fraud-detection model using XGBoost, with SHAP for explainability. Accuracy went from 62% to 73%, but the bigger win was that every flagged case came with a one-line explanation of which features drove the score.

Why it mattered

The officers stopped second-guessing, because they could finally see the evidence. That’s the point I keep coming back to: in production ML, the accuracy lift only matters if the people using it can act on it.

The model supports the humans making the call. It doesn’t replace their judgment, it shows its work so they can trust it.

Stack

Python, XGBoost, SHAP. Delivered through Simon Vision Consulting for the City of Rochester.

An academic partnership with an AI-focused finance firm, working under Professor Dan Keating from December 2024 to March 2025.

What I built

Four things, end to end:

• An AI-driven lead-nurture sequence
• A CRM tailored for insurance agencies
• A mini-site with AI features built in
• Brand and multi-channel marketing for SixPoint.AI and WerkWise.AI

What I took from it

Real AI applied to a real business, not a demo. Lead nurturing, CRM design, and AI-integrated sites are different problems, and shipping all four in one partnership sharpened how I move between data, system design, and the go-to-market side.

I worked alongside Ryan Lin and Professor Dan Keating, refining each deliverable and presenting the work to leadership.