Senior Product Manager · Toronto

Matt Hughes

Data-driven product veteran, twenty years into shipping streaming platforms, SAAS products and AI solutions. Lately I've been building small, playable proofs to understand agentic AI as well as I understand a roadmap.

[email protected] GitHub ↗ LinkedIn ↗ Picton, Ontario
See the experiments ↓ Download résumé ↓
Matt Hughes

Career so far

Started as a developer, spent most of the past two decades running product for streaming platforms, SAAS products and now build AI systems for clients at a boutique AI consultancy.

  1. 2022-Present
    Senior Product ManagerArchitech, Toronto

    AI solutions for a portfolio of clients ranging from startups to leading telecom and media companies.

  2. 2021-2022
    Senior Product ManagerProposify, Halifax

    Grew a contract management product using qualitative and quantitative data: integrations, A/B testing, workflow expedition, user interviews.

  3. 2020-2021
    Career BreakProfessional Development

    Modernized technical skills with certificates in React, Angular, Node.js, C, Python, R, Data Science, and Statistics.

  4. 2017-2020
    Senior Digital Product ManagerSOCAN, Toronto

    Led design and development of an API tracking third-party application results, integrating traffic analytics into development decisions.

  5. 2014-2017
    Manager, Product DevelopmentShomi.com, Toronto

    Designed core analytics and recommendation engines for a nationwide video streaming service.

  6. 2010-2014
    Product ManagerCBC.ca, Toronto

    Built a nationwide video streaming service centred on usage analysis and advertising technology.

  7. 2008-2010
    Flash DeveloperCanwest, Toronto

    Wrote ActionScript to implement Omniture analytics support into a nationwide streaming platform.

The AI Lab

Four live experiments in agentic AI, built to keep my product instincts current with what these systems can actually do: LangGraph state machines, RAG memory, multi-agent debate, tool-calling economies. Each one is playable, not just described.

The Philosopher's Bar app: thinker selection screen with a grid of philosopher portraits and a topic field
Experiment 01 · Multi-agent debate

The Philosopher's Bar

Put 2-4 historical minds in a room and see who breaks first.

A live debate engine: pick your thinkers and a topic, then choose the format (open Freeform, formal Oxford, or ratings-chasing Cable News) and watch the argument route itself, turn by turn, toward whoever the conversation has most activated.

Dynamic speaker selection Moderator steer breaks Consensus & tension detection Heat & concession tracking Live evidence injection Rolling context summarization

LangGraph·OpenAI GPT-4o / GPT-4o-mini·Tavily search

▶ Play live View code
AI Driven Haunted Mansion app: the study room scene with a character portrait, mood/fear meters, and the room's item map
Experiment 02 · Memory & social simulation

AI Driven Haunted Mansion

The ghosts remember what you told the butler.

A text adventure where every NPC keeps an independent memory, mood, and fear level. They gossip about you to the rest of the house the moment you leave the room, so a bad first impression follows you from study to cellar.

RAG memory (Pinecone + HyDE) Per-NPC emotional state Gossip network propagation Agentic shopkeeper ReAct web search Token-streamed responses

LangGraph·GPT-4o-mini·Pinecone

▶ Play live View code
Not Jeopardy app: player scoreboard and a category board with dollar-value clue tiles
Experiment 03 · Generative content & judging

Definitely Not an AI Rip-off of Jeopardy!

No question bank. The board writes itself.

Every category, clue, and answer is generated on demand (nothing pulled from a bank), then independently fact-checked before it ever reaches the board, and judged live against AI contestants with their own speed and accuracy quirks.

On-demand clue generation Independent fact-checking pass AI contestants with judging Host commentary + voice Full LLM tracing

GPT-4o·o4-mini·GPT-4o-mini·ElevenLabs TTS·LangSmith

▶ Play live View code
REPUBLIC OF VERIDIA POLITICAL SIMULATOR The country will not be improved by optimism.
Experiment 04 · Constrained simulation

Republic of Veridia

The economy doesn't care how confident your speech was.

A twelve-month political survival sim where deterministic game logic sets the hard bounds of truth and the LLM only narrates inside them, classifying faction reactions and generating flavour text, but never writing state directly. Four sub-stats drift on their own rules underneath every decision.

Engine-clamped LLM outputs Autonomous multi-stat economy Authored + AI-generated crises Scripted 26-beat campaign arc

LangGraph·OpenAI·FastAPI·React / Vite

▶ Play live View code