Bruce Krysiak — Venice, California
I see what’s possible before it’s obvious, then build the strategy, the team, and the product that make it real.
The scarce person on an AI team isn’t the researcher or the senior engineer. It’s whoever can see what the product could become, make the case for it, and then build the strategy, the organization, and the thing itself. Most people do one half or the other.
I’ve done that again and again, from both sides of the table. Sometimes it was my idea. Many times I was brought in to sharpen someone else’s. It is most of what I’m good at, and the only thing on this page I’d ask you to remember.
My master’s in EECS is from the MIT Media Lab, where I wrote a thesis on decentralized multi-agent systems. That was far ahead of the curve in 1995 and right on time today. Since then I’ve focused on the deployment side, where models meet real users and a P&L. The ability to build core AI engines is rare, and knowing what to build around them so the output is worth something is essential. I’ve done both, from scratch, in every generation along the way.
The Media Lab taught me the habit that matters most, a sensitivity to what is already possible but not yet visible. Work there runs a decade or two ahead of everyone else, and it gets there by noticing untapped value sitting in plain sight rather than by predicting anything. The most valuable thing I do is find the opening nobody has named yet, and ask the questions nobody thought to ask.
Multi-agent systems and constraint solvers in 1995. Collaborative filtering in 1998. Trust graphs in 2000. Embeddings and agents now. Not a career switch. It has been the same problem the whole way, systems that infer human intent and organizations that can ship them.
AI classes with Marvin Minsky. Undergrad work at Sloan. A drum machine built from the chip level up.
Media Lab thesis on decentralized control — multi-agent systems before anyone knew the term.
Constraint-solver platform behind enterprise configuration.
Collaborative filtering plus an expert rule layer, for music. I’ve never found an earlier one.
Trust graphs as decision infrastructure. Patent filed; 14 founding employees.
Contract abstraction — two decades before it became an AI category.
Predictive pricing within 5% of actual, across 600,000+ records.
Video aggregation and viral distribution. Consumer content dynamics at scale.
Intent inference from behavioral and crowdsourced signal. Acquired by TrueCar in 2010.
Machine + human-learned recommendations against the ad-blocking wave.
Privacy payments at scale. Roughly 22 engineers across four teams.
Agentic systems. Vector retrieval plus an expert rule layer. Structured extraction, offline resilience.
In 1998 I built the Genre Browser, possibly the first hybrid collaborative-filtering and expert-system engine for music discovery. I’ve never found an earlier one. Collaborative filtering represents people and music as vectors and recommends by proximity. It works, and it fails in specific, predictable ways: it collapses toward the popular, bridges incorrectly across semantic clusters, overindexes early data, can’t explain itself, and can’t tell when it’s wrong.
I’m also a DJ, so I knew things about why a recommendation lands that the statistics couldn’t see. I encoded that as a rule layer over the similarity engine, and it made the recommendations dramatically better.
Vector-space operations plus expert rules (aka skills and harnesses) is exactly where applied AI work sits right now. Embedding-and-retrieval systems are direct descendants of collaborative filtering: Magpie runs on pgvector for the same reason the Genre Browser ran on similarity. The practical job is still the same job. You find where the statistics break and you build domain knowledge into the structure around them: schemas, evals, constrained extraction, retrieval filters, and the interface that decides whether any of it reaches a person as something they want. Different vocabulary, same problem.
Every generation in AI has new strengths and weaknesses. The real job is: find them quickly, build the structure that compensates, then tear down and rebuild that structure when the engine improves, again and again. I’ve had thirty years of reps here.
From badly-specified mandates to specs, plans, and a shipping cadence.
Panjea: repositioned a social network fighting MySpace and Facebook, and shipped the new product in six months. TrueCar: absorbed an acquisition, migrated production, shipped the first mobile app. Markett: stabilized the platform in three months, launched the new app in six.
I build high-performing teams, and the processes that make them fast.
Aviri: recruited 14 founding employees. Accruent: 16 engineers onshore and offshore through a full platform rearchitecture. Honk: 140+ production releases on TDD and agile. MobileCoin: four teams, roughly 22 engineers; hired one engineering manager and promoted two more from within.
The engine is a third of the work. Harnessing with rules and guardrails corrects where the statistics fail, and the interface determines whether that value gets to users effectively.
1998: collaborative filtering plus a DJ’s rule layer, wrapped in a browsing experience for finding music worth buying. 2013: machine + human-learned recommendation, built for readers who were actively blocking ads. 2026: voice-first capture, because the fastest way to record what you know about someone is to say it out loud.
Independent project
2026 — present
Relationship intelligence for the agentic era. Voice-first capture, on-device Whisper transcription, structured LLM extraction, and pgvector semantic search across a FastAPI backend and a React Native app.
Private beta · ~4,200 commits in six monthsIndependent project
2025 — present
Any news article becomes a delivered message to your actual representatives. Underneath it, an independent address-to-representative discovery and delivery layer, and an open dataset covering the LA local-official tier that national civic data omits.
Live · ~950 commits across three reposDirector of Engineering
2021 — 2023
Ran the application layer of a $1B-valuation privacy-payments company — roughly 22 engineers across web, mobile, core platform, and an acquired bot team. Inherited a Ruby web group with no web-product track record and a departing PM, and carried it through a company-wide move to Kotlin: restarted the product, retrained the team, shipped in under three months. Also took over the team maintaining MobileCoin’s payment integration inside Signal’s mobile apps and kept it moving upstream into Signal’s own codebase.
Roughly 22 engineers · 4 teamsFounder / CTO, Venice Beach Labs
2013 — 2019
Machine + human-learned influencer recommendation, built against the ad-blocking wave. Multi-tier caching, high-throughput Instagram and blog crawling, Twilio SMS, a modular ES6 client, and JWT auth. Shipped under Channelverse, HAVN, and fAds.
Six years · three product brandsVP, Advanced Product Development
2010 — 2012
Arrived through the Honk.com acquisition with a mandate to integrate. Migrated production to EC2, shipped TrueCar’s first iPhone app, audited the legacy Java platform, and architected a platform strategy transition.
Acquisition integrationCo-founder / CTO
2008 — 2010
Vehicle discovery built on social signal and machine-learned recommendation. Alpha in one month, second alpha in two, private beta in three.
140+ production releases · Acquired by TrueCarCTO
2006 — 2007
Joined a social network for artists and creatives fighting MySpace, Friendster, and Facebook for the same ground. Repositioned the company to video aggregation and a widget platform that distributed creator channels across the web, installed QA and agile into a 16-person distributed team, and shipped in six months.
DEMOgod, DEMO 2007 · MIT TR35 nomineeTwo years of agentic prototyping and development, and two major lessons:
First, how much traditional software process carries over. Code review, tests, specs, small commits, staged rollout: all of it exists because people perform variably and you can’t just trust they got it right. That’s exactly the problem with an agent. That discipline transfers almost unchanged, and teams without it are learning that the expensive way, at 10x speed.
Driving and honing that process has been the key facet of my startup experience: how does a small team ship fast and still be right. Zero to one is where that question is most crucial, and it is the same question now, with tighter loops.
I built one in March. A Linear ticket dispatches an agent to plan, implement, review, and open the PR, with a triage gate and a kill switch over the top. People are calling that pattern a software factory now.
Then I mostly stopped using it. Working solo, iterating directly in chat is faster, and the pipeline had my own assumptions about task size baked into it. Those expire fast: as the models improve, the leaf nodes you can hand off whole keep getting bigger. Knowing when process earns its keep and when to tear it out is the actual skill.
Second, how much leverage is just about staying current. What works across models, skills and techniques shifts monthly. The compounding comes from tracking it, testing it, and adopting fast, not from any single tool choice. A team that stopped paying attention a month ago is already behind. A team that stopped paying attention six months ago lives in a completely different world.
People forget who they met and why, and teams miss the introduction a colleague could have made. Magpie captures it by voice in the moment, then makes it searchable. React Native, FastAPI, BAML, pgvector, on-device Whisper. Private beta on TestFlight.
People read the news, care, and do nothing. Not apathy: twenty minutes of friction sits between caring and actually making your voice heard. Article extraction, jurisdiction resolution, agentic drafting, delivery. Live.
LA fire victims learned they were underinsured at the worst possible moment. Policies are unreadable by design, and nobody reads one until they file. Upload, guided intake, coverage-gap analysis. Visual prototype.
Agent skills are executable code pulled from Git with no update path and no way to revoke a bad one. That’s a supply chain waiting to be attacked. Update and revocation framework, with a reference implementation.
Civic discourse is disintegrating under ragebait and dark engagement doom loops. A proposal for the revamp and rollout of the digital civic infrastructure a healthy democracy needs.
Streaming audio doesn’t hold up in a club, and hunting down high-quality versions of every track is tedious. Trackster upgrades a whole collection automatically, so the next set slams. In progress.
My instincts are founder instincts. I’m at my best uncovering a market insight and honing it into something that works as a business. That doesn’t mean I need to run the company, and it doesn’t make me hard to manage. I commit completely to a direction once I understand the reasoning behind it. What doesn’t work is being a middle manager relaying decisions I don’t understand and don’t agree with.
What I need is visibility more than authority. The opportunities I’m good at finding tend to sit between the pieces, in a recombination or an alignment nobody had reason to look for, so I have to be able to see the pieces upstream and down. Constraints drive innovation, and I can’t innovate against constraints I can’t see. Put me in the middle of a system I can only see one layer of and my judgment gets worse, because I can’t tell what an idea will collide with.
I can run a well-specified plan, and I’ve done plenty of it. But the value I add is almost always in the other case, taking something hard and badly specified and wrestling it into a shape a team can build against.
I want to work with great people, and I mean two things by that. Colleagues who are better than me at something that matters, who I’ll learn from. And people who are warm, who care, and who are trying to do right by whoever the work touches. I want to come into work excited about the people around me.
CTO, VP Engineering, or ownership of a major product surface at a smaller AI company where thirty years of pattern-matching compounds. Or applied research and lab leadership inside a larger one.
Combined BS/MEng in computer science and electrical engineering, completed in four years. Concentration in artificial intelligence, including classes with Marvin Minsky. UROP in Bill Dally’s parallel processing group. Undergraduate coursework at Sloan. Designed and built a drum machine from the chip level up. Minor in writing. Media Lab thesis: “Networked StarLogo — A Social Exploration of Decentralized Control Systems.”