Miami, FL · AI software engineer

Joshua Lazoff

I build AI that shows its work - or says nothing.

Twenty years of production engineering for names you know - Goldman Sachs, the team behind MetaMask, Chainlink. Now I build knowledge graphs and grounded AI: systems that name a real source for every answer, and stay silent when they cannot. Everything runs on hardware I own, measured by tests you can read and run yourself. I want to do this on the hardest problem you have.

The story

The story, in three acts

A kid fixing computers on an island. The job everyone wants, and the morning discipline nobody saw. A leap, a long apprenticeship, and then the year the machines arrived. The research thread running underneath has its own page.

  1. Act I · The island and the floor
  2. Growing up

    San Juan

    Puerto Rico

    Born and raised on the island. By high school I was running a 200-member youth organization - the activities, the budget and the fundraising, with a fifteen-person board. My first lesson in keeping a system running when every part of it is a volunteer.

  3. Engineering

    At fourteen

    Paying customers

    Building and programming computers for people in San Juan who paid me for them.

  4. Engineering

    Four years, then one more

    Cornell

    Engineering, then a master's. Information science is the study of how to represent the physical world in a digital one, which turns out to be a fair description of everything I have done since. I taught calculus to other students and kept the books for the Hispanic engineers' chapter.

  5. Engineering

    Three summers running

    Goldman Sachs, as an intern

    The same trade-processing team three years in a row, building the tools that check the firm's trade records against its own risk system. Somewhere in there I also started running the mentoring program for the incoming interns.

  6. Engineering

    Then the job itself

    The trading floor

    Software engineer in Securities technology, on the systems that double-check trades and file the firm's regulatory reports. I built a data-lookup library the whole firm could use, and helped stand up a round-the-clock operation that catches trade errors and fixes them.

  7. Act II · The leap, and the long apprenticeship
  8. June 2015

    Two names, one month

    I gave up the “perfect job.” That same month I started two companies - LazoffTech for client engineering, and a second for a long-running personal research project: that story has its own page. The first has funded the second ever since.

  9. Research

    The next two years

    Learning platforms

    Senior engineer on a structured-learning platform built with a team in Israel, then a ground-up rebuild of a video-learning platform. Two years of shipping products where the content model - not the code - was the hard part.

  10. Engineering

    Six years

    ConsenSys

    I came in through the education arm, building its online learning platform, and stayed six years. I left as the senior engineering manager for every website behind metamask.io and consensys.io.

  11. Engineering

    Three and a half years

    Chainlink Labs

    Contract engineer inside Chainlink - a company that feeds real-world data like prices and events onto blockchains - building the sites and tools its developers rely on. The contract ran through TorahTech LLC: the blockchain income flowed through the Torah company.

  12. Act III · The machines arrive
  13. Engineering

    2023

    A machine in the house

    Everything the career taught - connecting information, making it findable, running systems that do not lose things - pointed at one product: private AI on hardware you own, sitting on your own shelf.

  14. Engineering

    2026

    The year of the machines

    When the AI tools matured, everything I had spent two decades saving became fuel. In one year I shipped more new projects than in every earlier year combined - sites, graphs and agents, built and run from a fleet of machines in my own house. One person, working like a team.

  15. Research

    Now

    OpenRabbi, my grounding testbed

    A knowledge graph of a classical text corpus, every piece carrying its source - with eight conversational voices on top, each one grounded in the graph. The generation rule is the point: cite a real, checkable source, or say nothing at all. It runs in public, on hardware in my house.

  16. The rule, measured

    Does the rule hold?

    It holds. The whole system turns on one constraint, written into the code.

    Every answer must name a real source you can go and look up, verified word for word against the actual text - and if no source exists, the system says nothing. Only a model that clears every check ships. The rule is the oldest one on the shelf, and it was always the rule: say who said it, or say nothing.

    That discipline - cite or stay silent, measured, in public - is what I want to bring to a team working on AI you have to be able to trust. The research behind it is at jew.tech; the productized version is Digital Twin Pro.

What I build

Systems that have to show their work

One rule runs through everything: an answer arrives with the source it came from, or it does not arrive. Everything below runs today, in public, on hardware I own.

Knowledge graphs · grounded AI

A graph that answers with its sources

A 157,000-piece knowledge graph of a classical text corpus, with conversational AI on top. Inside it, a citation network of 2,185 authors joined by 2,504 teacher-to-student links. Ask it anything and the answer arrives with a reference you can look up - then you can go check it.

Live now as OpenRabbi - a speaking agent grounded in the graph, running from my house.

Evaluation · testing

Every citation verified, word for word

Every citation is verified word for word against the real text before anything ships, in written tests and again in live conversation. You can read the methodology, run the queries behind every number, and check any figure yourself.

Infrastructure · product

A fleet, run like production

Forty-nine live sites built, deployed and QA'd by pipelines I wrote, on a GPU cluster in my house - and Digital Twin Pro, the company I founded: a private AI appliance you buy once and own outright.

Underneath all of it: machines in my own house. A system built to show you where its answers came from should be one you can open up and look inside. The personal research this grew out of is at /torah/.

Built and running

The sites

49 live sites. Every one of them started as something I wanted to understand well enough to explain out loud. Here are six to start with; the rest, with what each one is for, live at jew.tech.

Connect

One conversation I always take

Are you building AI that has to be trusted before it is believed? Do you need someone who ships, measures, and publishes the receipts? Do you want an engineer who has run this exact discipline in production for years, not proposed it in a slide?

That is the conversation I always take. Write to me and I will answer - with sources!

Want the full picture first? The research is at jew.tech, the receipts at jew.tech/evidence, and the experiments I want to run next at jew.tech/agenda.