Product manager · Logistics & AI

Carl Jean-Louis.

Product manager who builds. I’m a product manager at UPS, modernizing the legacy COBOL platform behind customs filing for roughly 200 countries — onto SQL and Google Cloud — and automating the filing process itself. Outside the day job I take AI products the whole way myself: problem definition, market analysis, spec, architecture, evaluation, and production deploy. The work below is that second half.

Experience Email GitHub
~200
countries in scope at UPS
5 yrs
at UPS, 1 as product manager
18
products shipped solo
15
live in production

Experience

Where the product work happens with other people, on a roadmap I don’t unilaterally own. Legacy modernization and customs automation at global scale.

UPS

Product Manager

Aug 2025 — present

Modernizing the legacy platform behind customs filing for roughly 200 countries — and automating the filing process that runs on top of it.

UPS

Special Project Coordinator

2021 — 2025

Deploying new operational technology across the delivery network.

How I decide

Six product calls I’d defend in a room, each one from work I actually shipped — including the two where the evidence went against me.

01

Pick the metric that maps to money, not the one that flatters the product

Scoping an AI system for logistics exception handling, I made the headline metric false automation rate — the odds it autonomously does something consequential and wrong with no human review — instead of accuracy. Accuracy conflates that with the harmless failure of escalating something it could have handled, and those two errors cost wildly different amounts. I always pair it with autonomous coverage, because false-automation rate alone is trivially gamed by escalating everything.

02

Design the experiment so your own design can lose

I included the control condition most multi-agent comparisons quietly omit — a single model sampled repeatedly with a majority vote — on the explicit understanding that if my specialized architecture could not beat it, the architecture was ceremony and the write-up should say so. The control won. The write-up says so in the third paragraph, and the recommended design became the boring one.

03

Lead with the bear case

My strategy docs open with the strongest honest argument against building the thing — named competitors, their funding, their delivery model, and the precedent where this category already disappointed once. A strategy document that buries its own disconfirming evidence is a sales deck. On the same project I scored nine candidate category names and rejected the one I had proposed, because it routed every deal to the wrong budget holder.

04

Complexity has to earn its place

In a content intelligence product, a short essay gets a four-node pipeline; a ninety-minute podcast dense with factual claims gets parallel claim researchers and bounded research loops. Same product, cost scaled to the input. Related call: research runs once and produces one packet, and every output format is a rendering of that packet — asking for four formats does not run four researches. That decision lives in the data model, where it is cheap, rather than in the feature, where it is not.

05

Ship the uncertainty instead of polishing it away

Every claim my research system produces carries an epistemic status — verified, disputed, refuted, speculation, opinion, unresolved — and the outputs are structurally forbidden from blurring them. A claim that survives two rounds of research unresolved gets reported as "I could not independently verify this," including in generated audio, where smoothing it over would have been easy and invisible.

06

Find the load-bearing feature by testing, not by intuition

In a game I built, upgrades started as a nice-to-have. Playtesting showed the economy broke without them — currency piled up unspent once build slots ran out and defense could not scale. So I encoded balance as assertions: every map must be clearable, an undefended base must always lose, and no single tower type may clear the hardest map. Balance is tested, not felt. The same instinct applies to product: state the acceptance criteria precisely enough that they can fail.

Product craft

  • Requirements gathering
  • Epics & feature decomposition
  • Acceptance criteria
  • Backlog ownership
  • Prioritization & cut decisions
  • Stakeholder alignment

Strategy

  • Competitive analysis
  • Positioning
  • ICP definition
  • Buying-committee mapping
  • Phase-gated roadmaps
  • Build / don’t-build calls

Metrics & evaluation

  • Metric selection
  • Counter-metrics
  • Falsifiable experiments
  • Control conditions
  • Stating what data cannot support

AI product

  • Model routing & cost tiering
  • Agent system design
  • RAG & vector search
  • Evaluation harnesses
  • Prompt-injection threat modeling
  • Speech & image pipelines

Domain

  • Global logistics & supply chain
  • Customs & cross-border compliance
  • Exception handling, OS&D, detention
  • Regulated, multi-country environments

The Index

18 products designed and shipped end to end between 2025 and 2026 — spec, build, and deploy. Filter by what you’re curious about.

Conductor — A PM command center that turns meetings into knowledge. 01
AI Products

Conductor

A PM command center that turns meetings into knowledge.

Conductor transforms meeting recordings, documents, and notes into an intelligent, searchable knowledge base. It transcribes audio with speaker diarization, auto-extracts summaries, action items, and decisions, and includes a Q&A chat plus a writing assistant that answers in your own voice and references your content.

  • Cloudflare Workers
  • D1
  • React
  • Vector search
NimbleLogic AI — AI automation studio for real businesses. 02 Live product
AI Products

NimbleLogic AI

AI automation studio for real businesses.

NimbleLogic builds AI automation for any business with a receptionist or back office — AI phone agents that answer and book calls, AI personal assistants, and back-office operations automation. It serves verticals from legal and healthcare to real estate, salons, and contractors.

  • Next.js 14
  • Cloudflare Pages
  • Voice AI
  • Telephony
VoiceClaw — Real-time speech-to-speech with an AI agent. 03
AI Products

VoiceClaw

Real-time speech-to-speech with an AI agent.

VoiceClaw is a real-time voice interface for continuous speech-to-speech conversations with an AI agent. It uses LLM streaming with sentence buffering and cross-channel context syncing to hold latency under ~1.5 seconds, with a mobile-optimized UI.

  • Cloudflare Pages
  • FastAPI
  • Cloudflare Tunnel
  • Streaming LLM
ForgeStory — Turn any text into illustrated visual stories. 04
AI Products

ForgeStory

Turn any text into illustrated visual stories.

ForgeStory turns any text into a visual story — AI generates illustrations and narration so reading becomes something you watch. Built on the idea that learning should be visual whether you’re 8 or 80, with a personal library, accounts, and subscriptions.

  • SvelteKit
  • Cloudflare Workers
  • Image generation
  • TTS
Echo Studio — A personal AI media studio on the edge. 05
AI Products

Echo Studio

A personal AI media studio on the edge.

Echo Studio converts text to natural-sounding speech, generates AI podcasts that combine live web research with AI scriptwriting, and transcribes audio with Whisper — all running on Cloudflare’s edge network with a polished, mobile-first dark UI.

  • React
  • Cloudflare Pages
  • Whisper
  • Claude
Tribunal — AI models debate, so you see every side. 06
AI Products

Tribunal

AI models debate, so you see every side.

Tribunal stress-tests an idea by having multiple AI models argue it from opposing positions and then reach a verdict. It’s a tool for pressure-testing decisions and surfacing the strongest version of every argument before you commit.

  • React
  • Vite
  • Cloudflare Workers
  • Multi-model LLM
Atlas Library — A personal hub for long-form text and audio. 07
Voice & Audio

Atlas Library

A personal hub for long-form text and audio.

Atlas Library is a personal content hub for long-form writing and audio with AI-assisted reading. Articles, briefs, and deep dives are stored, narrated, and made listenable — designed for an audio-first reading habit.

  • TypeScript
  • Cloudflare Pages
  • TTS
French Transcriber — French speech-to-text with Whisper. 08
Voice & Audio

French Transcriber

French speech-to-text with Whisper.

A focused French speech-to-text transcription app powered by OpenAI Whisper — upload audio and get clean, accurate French transcripts.

  • Whisper
  • JavaScript
  • Cloudflare Pages
BeatForge — Create. Play. Forge your sound. 09 Demo offline
Voice & Audio

BeatForge

Create. Play. Forge your sound.

BeatForge is a browser-based beat maker and music creation studio — sequence, play, and craft sounds directly in the browser, backed by a Cloudflare Workers API.

  • React
  • TypeScript
  • Web Audio
  • Cloudflare Workers
VoiceGen — AI voice generation. 10
Voice & Audio

VoiceGen

AI voice generation.

VoiceGen (VoiceForge) is an AI voice generation app for synthesizing natural speech from text with selectable voices.

  • TypeScript
  • TTS
Atlas Daily — An automated morning intelligence brief. 11
Automation & Intelligence

Atlas Daily

An automated morning intelligence brief.

Atlas Daily is an automated daily briefing system — it scans overnight news, digests email, and assembles an AI-curated morning brief tuned to a specific set of interests, delivered as audio for the commute.

  • TypeScript
  • Cloudflare
  • LLM agents
  • TTS
Atlas AI Radar — AI news, automatically turned into short videos. 12 Pipeline
Automation & Intelligence

Atlas AI Radar

AI news, automatically turned into short videos.

A fully automated pipeline that turns AI news into short-form videos: OpenAI for images, fal.ai Seedance for motion, ElevenLabs for voiceover, and Remotion/FFmpeg for assembly — an end-to-end content engine running on Cloudflare.

  • Python
  • fal.ai Seedance
  • ElevenLabs
  • Remotion
Atlas Hub — A unified personal life dashboard. 13 Private
Automation & Intelligence

Atlas Hub

A unified personal life dashboard.

Atlas Hub is a private personal dashboard that unifies wealth, health, and memory — real-time banking and investment tracking, Fitbit health trends, a synced daily memory feed, and a secure document vault. Kept private; described here for context.

  • SvelteKit
  • Cloudflare
  • D1
  • Plaid
Sentinel — Prediction-market monitoring & paper trading. 14 Private
Automation & Intelligence

Sentinel

Prediction-market monitoring & paper trading.

Sentinel is a Polymarket monitoring and paper-trading dashboard — it tracks markets and simulates trading strategies. Kept private; described here for context.

  • SvelteKit
  • Cloudflare
  • Polymarket API
Duo — Couples goal & habit accountability, with an AI coach. 15
Apps & Experiments

Duo

Couples goal & habit accountability, with an AI coach.

Duo helps partners track shared goals and habits, plan adventures together, and stay accountable without nagging. An AI coach offers personalized insights, goal breakdowns, and weekly progress summaries.

  • Next.js 14
  • Cloudflare D1
  • Tailwind
  • Claude
Chess + AI Coach — Play Stockfish with a Gemini coach explaining every move. 16
Apps & Experiments

Chess + AI Coach

Play Stockfish with a Gemini coach explaining every move.

A browser chess game backed by the Stockfish engine, paired with a Gemini-powered AI coach that explains the position, suggests ideas, and teaches as you play.

  • React
  • Vite
  • Stockfish
  • Gemini
Convergence — Run an AI lab and race to AGI — a strategy sim. 17
Apps & Experiments

Convergence

Run an AI lab and race to AGI — a strategy sim.

Convergence is an AI research-lab strategy simulation: manage compute, talent, safety, and competitors as you race toward AGI. A playable take on the real dynamics shaping frontier AI.

  • JavaScript
  • Game design
  • Cloudflare Pages
LemonWater — A hydration tracker with daily goals. 18
Apps & Experiments

LemonWater

A hydration tracker with daily goals.

LemonWater is a simple, friendly hydration tracker — log your water and lemon-water intake against daily goals, with accounts to keep your streak going.

  • SvelteKit
  • Cloudflare
  • Auth