Product management leader with 15+ years in B2B SaaS across fintech and healthtech — building agentic AI systems end-to-end. Not prototypes. Not slideware. Real builds running on real infrastructure, demonstrating what a PM AI builder looks like when the build barrier collapses.
Each agentic system solves a non-trivial coordination problem — multi-agent orchestration, adversarial reasoning, cross-source signal fusion, human-in-the-loop gating, persistent memory loops. Built solo. Production-deployed. Each a complete proof point.
A pre-build feasibility simulator for AI agent products. A product manager describes a proposed agent in 8 questions. Twenty minutes later, they walk into the kickoff with a confidence score and the three risks most likely to kill the project — each with a specific mitigation move.
The pattern today is to start building an agent on a hunch and discover three weeks in that the eval signal is too noisy to ship. The wasted sprint costs $80K–$200K loaded, plus credibility. Gartner forecasts 40%+ of agentic AI projects will be cancelled by end of 2027. Conductor PM attacks the decision that precedes all three. A scan of fifteen agent-eval platforms — LangSmith, Braintrust, Arize, Maxim, Coval — confirms the wedge: every one scores a built agent. None scores a described one. Three probes run in parallel: eval-signal noise, tool-call composability, failure-mode tolerance. Output: 0–100 score + 3 ranked risks + prescriptive GO / PAUSE / KILL.
The differentiator: A PM-native buyer at a different moment than the eval incumbents. They sell to engineers after the build. Conductor PM sells to product leaders before the spend — and tells them what their team would otherwise learn in three weeks.
Most "AI for marketing" tools generate content. Conductor manages the work. It plugs into Linear — the tool engineering teams already trust to track tasks — and turns it into a control panel for AI agents that do the actual job: drafting campaigns, refreshing copy, running competitive scans — gated by human approval at every step.
The breakthrough is the control panel. Engineering teams already use issue trackers like Linear to assign work, track progress, and approve completion. Conductor uses that same workflow — but the "worker" is an AI agent, and the "manager" is the human in charge. Each ticket becomes an autonomous task: research, draft, refine, hand back for review. Built for the work most companies aren't automating well yet — marketing ops, RevOps, lifecycle, customer experience. Open-source, runs on Claude, 14-day build.
The differentiator: Conductor is built on top of OpenAI's Symphony pattern — originally designed for engineering workflows — and extends that same approval-gated control panel into marketing, RevOps, and lifecycle teams.
Reconciliation for agentic payments. As autonomous agents transact across cards, stablecoins, and ACH, the reconciliation surface fragments. Throughline catches the exceptions a finance team can't afford to miss — amount mismatches, duplicate settlements, over-threshold authorizations — across every rail.
Built on direct payments expertise — $3B+ in embedded payments volume at Planet DDS, plus payment-orchestration architecture experience. The premise: when AI agents start moving money on their own, the gaps that humans used to catch in spreadsheet reconciliations become structural. Throughline sits post-settlement and detects three exception types — amount mismatch, over-threshold authorization, duplicate settlement — across card and stablecoin rails uniformly. Built as a portfolio piece to demonstrate domain expertise meeting agentic commerce.
The differentiator: Most "AI for payments" pitches generate marketing copy. Throughline solves the actual operations problem agentic commerce creates — reconciliation surface fragmentation as agents move money across rails — using the same finance-grade rigor incumbents apply to human-initiated transactions.
Autonomous multi-agent product pipeline that ideates, debates, plans, and builds a new product daily — without me.
Daily 6 AM autonomous runs via Mac LaunchAgent: Discovery → CEO Debate → Gate 1 → 7 Planning agents → Gate 2 → Build → Gate 3. Three HITL approvals through Telegram. CEO Debate Layer surfaces risk through opposed PM/CTO/CFO/Marketing/Legal agents before the synthesizer renders GO / NO-GO / NEEDS_REVIEW.
Agentic stock-debate orchestrator. Bull / Bear / Contrarian agents reason adversarially over structured evidence bundles, daily.
Three Sonnet agents on Claude Code Agent Teams, restricted to messaging primitives — no tools, no web — debating over a SignalBundle assembled from Databento market state, SEC EDGAR filings, Alpha Vantage news, and Finnhub catalysts. Completeness gate: bundles below 90% don't go to debate. First live run validated the architecture — Bear caught Bull on incorrect base-rate arithmetic; Bull conceded. The disagreement is the product.
Three-signal convergence detector — fuses derivatives flow, Polymarket odds, and news-silence windows to surface pre-announcement signals.
Bloomberg fuses this for $24K/year. Nothing under $5K does. Convergence Scanner watches for anomalous order flow in a macro-sensitive instrument, a concurrent Polymarket odds shift on a related event, and silence on the news wire — and fires a Telegram alert with a Claude-generated hypothesis when all three align. Not a front-running tool; a tailing tool on public order flow you can see.
Full materials pipeline on top of Claude Managed Agents — and the same system I use to land the interviews this portfolio supports.
V1 used a Managed Agent to search jobs (slow, fragile, 9 production bugs). V2 collapses that layer to SerpAPI + Haiku scoring (5–8 sec end to end) and reserves Managed Agents for what actually needs reasoning — a 5-tab materials pipeline: Resume Tailor, LinkedIn Outreach, Cover Letter, Company Intel (live web search), Interview Prep. A Grader agent writes to Turso memory; future runs read those signals back into prompts. Adaptive without overcorrection.
Working prototype of Google's Agent Payments Protocol — a coordinated multi-agent commerce flow with the full mandate chain end to end.
User intent ($700, Palm Springs, first weekend of November) → prime agent negotiates simultaneously with airline + hotel sub-agents → dual cryptographically-linked Cart Mandates execute in one atomic transaction → flight + hotel receipts written together or rolled back together. A 4-agent topology (Gemini Pro prime + three Flash sub-agents) at 64% cost reduction vs all-Pro without quality loss. The coordinated_booking_id is the proof both receipts came from one Cart Mandate.
7 AM ET pre-market scanner that turns the SEC EDGAR firehose into a ranked, Claude-scored Telegram alert before the open.
Pulls the prior ~18 hours of 8-K filings, scores them by Item type (1.01 Material Agreement, 2.01 Acquisition, 2.02 Earnings, 5.02 Officer changes) and catalyst keyword overlap (FDA, DOD, hyperscaler counterparties, rare earths, design wins, scale terms like "billion"), then posts a ranked alert via Telegram with the ticker, the excerpt, and a directional bullish/bearish call. Runs as a Mac LaunchAgent alongside ProdAgentCo — same operational pattern, different signal.
Every project here is deployed, tested, and reachable. Each one started as a question — "could a PM build this in a weekend?" — and ended with a live URL.
Vertical-branched 7-dimension assessment for executive prospects. Generates personalized readiness reports with engagement tier recommendation, preliminary agent strategy, and phased project roadmap. Doubles as ProdAgentCo's proposal-generator.
Next.js · Supabase · Tailwind · Vercel
Claude Sonnet 4.6 — personalized narrative reports need real reasoning; Haiku was too flat for executive output
Daily AI news dashboard built primarily for my 81-year-old father. RSS-fed for cost, Claude Haiku for synthesis, on-demand loading. Includes a dedicated Stillwater Ponies high-school sports tab as the highest-priority feed. Runs at ~$2/month.
React TS · Express · RSS · Vercel
Claude Haiku 4.5 — light news synthesis on a $2/month budget; reasoning depth would be wasted here
Ambient clinical documentation app. Browser MediaRecorder streams to Deepgram nova-2-medical for real-time transcription; Claude generates structured SOAP notes from the transcript. First production AI build — and a clear domain-to-AI bridge for healthtech roles.
Next.js · Deepgram nova-2-medical · Vercel
Claude Sonnet 4.6 for SOAP note structuring — Haiku missed clinical nuance in early tests
Real-time Polymarket whale-activity dashboard. WebSocket connection to the CLOB with 30-sec REST gap-filler, 3-layer dedup, NegRisk detection, and a 4-panel UI (live feed · top markets · volume spikes · whale wallets). Built and deployed in a single session.
Python FastAPI · Next.js 14 · SQLite · Telegram · Railway
No LLM in the hot path — pure WebSocket trade detection. Claude Haiku 4.5 only for alert summaries when fired
3-agent prediction chain for NBA, NHL, and NCAAM — model probability, edge gates, fractional Kelly sizing, Brier-score grading. v3.0 caught a 5-play NHL slate with +10.2% top edge. Originally a Replit build; later wired into a MiroFish multi-agent simulation seed pipeline.
Flask · SQLite · Replit
Claude Sonnet 4.6 for the 3-agent prediction chain — adversarial reasoning matters for edge calls
A book-direct vacation rental site at johnsonlake-home.com — full SEO meta stack, Open Graph cards, tabbed gallery with lightbox, embedded map, staging branch workflow on Vercel. Optimized to win hyper-local searches the big OTAs don't bother indexing for.
HTML · CSS · Vanilla JS · Vercel
PRD v1.2 baseball picks model — MLB Stats API, pybaseball, The Odds API, and Open-Meteo for weather signals. Adaptive feedback loop logs outcomes and adjusts signal weights, with locked-signal rules to prevent overcorrection on noise.
Python · MLB Stats API · pybaseball · Open-Meteo
Claude Sonnet 4.6 for signal-weight reasoning · Haiku for fast feature extraction across game logs
For 15 years I've led product teams in fintech, healthtech, and dental practice management — including $25M+ digital-lending portfolios at Equifax, all RCM product at Planet DDS ($10M ARR with $2M new ARR from AI insurance verification), and embedded payments operating at $3B+ in annual volume. Before that, two US patents from Canon R&D in NLP and augmented reality.
What changed in the past year is the floor. The build barrier has collapsed. With Claude Code, Vercel, and a coherent architecture, a senior PM can ship a multi-agent system in days that would have required a team a quarter ago. The bottleneck moved — from "can you build it" to "did you design the right thing."
That's the wedge I'm building into. Domain expertise as the defensible moat, the build itself as the proof. Every project in this portfolio shipped because I designed the architecture, wrote the spec, made the model and protocol choices, and ran the integration loop end-to-end. Not vibe-coded. Designed.
I'm currently in conversations for senior product leadership roles in B2B SaaS — where deep product fundamentals and hands-on AI fluency both matter. If you're building something that needs a PM who can hold the technical depth and the customer truth in the same head — let's talk.
Six lessons compounded across ProdAgentCo, Tribunal, Convergence Scanner, Edge Scanner, and the materials pipeline. These show up in interviews more than the resume bullets do.
Bull/Bear/Risk in Tribunal and the CEO Debate Layer in ProdAgentCo both proved it. It's not about more agents — it's about opposed agents. Forced rebuttals surface what a single agent rationalizes away.
Haiku for fetch / structure / filter. Sonnet (or Opus) for genuine reasoning and synthesis. The Advisor pattern delivers major cost reduction without touching output quality — once the workload is understood.
A 4-dimension matrix — irreversibility, blast radius, compliance exposure, confidence — gives a defensible answer to "should the agent act autonomously here?" Approve / edit / reject beats binary go/no-go.
Each agent owns one clearly defined responsibility. Overlap creates ambiguity that cascades through the entire pipeline. Define what each agent owns before writing a single prompt.
Agents that log outcomes and adjust signal weights are categorically different from static ones. But: explicit rules for how many outcomes trigger adjustment, and which signals are locked — or you overcorrect on noise.
CrewAI was the right starting point. LangGraph's interrupt() + checkpointing is becoming the production HITL standard. Build to migrate, not to last forever. The protocol layer is where the real durability lives — MCP, AP2, A2A.
15+ years of B2B SaaS product leadership across fintech, healthtech, and dental practice management — owning $25M+ portfolios, shipping AI in production, and now building agentic systems end-to-end. Click through, or download the full Word doc.
Product leader with 15+ years scaling B2B SaaS platforms across fintech and health tech — owning $25M+ portfolios and leading high-performing PM teams from strategy through execution. Proven track record shipping AI-powered products in production, with hands-on fluency across LLM pipelines, multi-agent orchestration, and full-stack AI product delivery.
If you're hiring a product leader for B2B SaaS, reach out. Fintech, healthtech, payments, and beyond — happy to discuss roles where deep product fundamentals and hands-on AI fluency both matter.