AI PRODUCT + SYSTEMS ENGINEER
I help founders turn ambitious, incomplete AI ideas into working products—especially when the hard part crosses product decisions, software, and the real world.
See the work →This was not a planned sequence. Looking back, each project forced a different answer to what software could infer, what it could safely do, and what should count as evidence afterward.
Cerebralx / Xone put me inside electrodes, noisy EEG, firmware, embedded hardware, and the gap between a compelling demonstration and a dependable loop.
I learned to treat measurement, feedback, and failure as part of the product—not implementation detail.PerPitch and perMail were independent products in different contexts. Both made timing, ambiguity, and the handoff after a conversation impossible to ignore.
I became less interested in a model sounding capable and more interested in the work the conversation could reliably complete.Julie and CarbonCompute approached machine action from different directions: enterprise records in one, funded human work in the other.
Both reinforced the same boundary: interpretation can be probabilistic; authority, state, and settlement cannot be vague.Each chapter shows one recognizable product moment, what I owned, and the strongest public evidence the record supports.
A phone agent that knows when the call needs a human.
An AI receptionist that handles missed calls, identifies urgency, and turns the conversation into a clear owner handoff.
Sole builder across voice, backend, and handoff.
Voice loop, backend, handoff, alerts, and reliability debugging.
Customer identities, numbers, recordings, and caller data remain private.Turning messy meeting notes into safe CRM actions.
Meeting notes become reviewable CRM proposals, approved writes, and verified system-of-record state.
Product and system design across the governed workflow.
Acceptance evidence covers successful ordered execution and duplicate-approval reuse.
Workflow evidence, not customer adoption.“Move the Acme follow-up to Friday and assign it to Maya.”
A founder should finish the thought before the AI performs intelligence.
A voice-first investor simulation that protects the founder's turn while deeper analysis happens outside the live exchange.
Founder and product/AI lead.
Recorded product-use context across PerPitch's founder-evaluation work.
Session evidence, not paid-customer, fundraising, or institutional-adoption evidence.When an agent commissions human work, payment should follow proof—not trust.
A testnet task marketplace where agent-funded work settles only after explicit verification or dispute.
Product, contracts, verification lifecycle, and frontend.
The documented LaborEscrow contract is inspectable on Ethereum's Sepolia testnet.
A testnet deployment, not mainnet usage, production funds, or protocol compliance certification.Turning visual attention into a command a connected device can understand.
A lower-cost BCI prototype mapping visual-attention EEG from a fixed target set to predefined device commands.
Founder and principal engineer.
Government-backed support associated with the Cerebralx venture period.
Funding support for venture development, not product revenue, sales, or clinical adoption.The model is one layer. The product still has to decide what may happen, who controls it, and how the result becomes visible.
Name what the model may infer, what the product already knows, and what must remain unresolved.
Interpretation boundaryTurn consequential behavior into narrow actions with explicit owners, permissions, and approval points.
Action boundaryRead back durable state, expose partial failure, and leave enough evidence to recover or reconcile.
Verification boundaryI work directly with founders and product leads when the brief is still moving and the product needs judgment across interface, model behavior, backend state, and delivery.
Expect short feedback loops, working software early, written decisions, and direct conversations about what the evidence does—or does not—support.
We identify the user moment, system boundary, and assumption most likely to invalidate the product before polishing the easy parts.
A decision map and a testable first sliceI work across product and implementation to make the live call, model handoff, external write, or device loop real enough to judge.
Working software around the hardest pathWe add approval, readback, audit state, and recovery where the product needs to earn trust rather than merely announce success.
A system the team can evaluate and continueProduct and system audits · focused risky-slice builds · embedded 0→1 delivery
Tell me where the product is stuck →I’m extending applied product work into the technical foundations that can justify trust: task-based evaluation, constrained deployment, and stateful agency.
Testing whether an agent completed the real task—not whether its explanation sounded convincing.
Smaller and on-device models where privacy, latency, cost, and hardware are product constraints.
How agents maintain state, understand an environment, and connect reasoning to grounded action.
Tell me what you are trying to solve, what exists today, and where it breaks. A few honest sentences are enough to begin.
Email me the rough version →