Most AI programs don't stall on difficulty. They stall on ambiguity.
The Catalyst Group is a technology and strategy club at UC Berkeley. Four leads who have shipped software inside companies and run experiments inside labs. We take one narrow question at a time and answer it with running code.
Enterprises have more AI ideas than capacity to test them.
Every large company already has a list of AI ideas. The list is not the problem. Most of the items on it could be built by a competent team in a few weeks.
The problem is the decision. Which of them deserve real budget, real headcount, and a place on next year's roadmap? That call needs evidence, the evidence needs a prototype, and no internal team has the capacity to prototype them all. So the list waits. A deck gets made about the list. Two quarters later the same ideas come back better formatted and no better understood, and by then the research has moved and half the assumptions are stale.
Closing that loop takes three skills in one room: someone who can read the current literature, someone who can write code that survives contact with real data, and someone who has stood at a bench and knows what the instrument actually outputs. Inside one org chart that combination is rare and expensive. At a research university it is ordinary. Between the four of us we cover it, and we can turn a question into a working answer in weeks rather than quarters.
That is the entire reason this club exists.
Operating premise
Building the prototype is the cheap part. Knowing which prototype to build is the whole job.
We are students. That is the trade.
Being a student team is a real advantage in some places and a real limitation in others. Both are below, so you can decide quickly.
What you get
The people you meet are the people who commit.
Four leads, all technical. There is no layer between the scoping call and the pull request.
Weeks, not quarters.
One narrow question, answered inside a semester and usually well inside it. Weekly demos make the pace visible instead of asserted.
Research read first-hand.
One of us does AI research at MIT CSAIL; another runs experiments at Berkeley Lab and the Innovative Genomics Institute. You get the translation, not a summary of a summary.
A student rate.
Materially less than a consultancy would quote for the same scope. That is part of the pitch and we will not pretend otherwise.
Direct access to Berkeley.
Labs, faculty, and the people running the instruments are a walk across campus rather than a procurement cycle.
What we are not
Not a strategy firm.
No org design, no change management, no hundred-slide readout. If that is the need, hire the firm; we will say so on the first call.
Not a staffing contract.
We take one problem at a time and finish it. We are not a bench you can draw against for a year.
Not unlimited.
We have coursework. We scope around it, state our real capacity upfront, and would rather decline than overcommit.
Not a dependency.
You own the repository and can run it without us. Documentation is written along the way, not the week we leave.
Not in the business of flattering a premise.
If the idea does not hold up, the deliverable is a clear account of why. That is still a result worth paying for.
Why the address matters
Berkeley is not a credential we borrow. It is the reason the literature, the instruments, and the people who work on both sit within walking distance of each other.
Three of the four leads are in M.E.T., a dual degree in engineering and business. The fourth is in EECS and Applied Math. In practice that means the person writing the evaluation harness has also sat through the build-or-buy argument, so neither conversation needs translating for the other. It also means a question about an assay, a chromatogram, or a genome-editing workflow gets answered by someone who has run one.
M.E.T.
Management, Entrepreneurship & Technology: engineering and business as a single dual degree, taken together rather than in sequence. Three of the four leads.
EECS
Berkeley's electrical engineering and computer sciences department, where the systems, applied math, and machine learning training comes from.
Berkeley Lab · JBEI
Bench research at the Joint BioEnergy Institute, up the hill from campus, on instruments that produce the messy data everyone else models.
Innovative Genomics Institute
Genome-engineering work on campus, and first-hand exposure to the protocols and constraints that scientific software has to respect.
The Bay Area
The companies deploying this work are a train ride away, which shortens every feedback loop that matters.
The Catalyst Group is a student organization at the University of California, Berkeley. This site is not an official publication of the University of California, and nothing on it should be read as a statement by the University, its national laboratories, or its faculty. Institution names appear here only to describe where our members have studied and worked.
Five rules we don't negotiate
Projects like this fail in predictable ways. These are the cheapest fixes we have found, and we apply them from the first week.
One problem at a time.
A single question per project, written down in one sentence. Breadth is the main reason work like this never finishes.
A success metric before any code.
We agree in writing on what would make the answer a yes, and what would make it a no. If we cannot write that down together, we are not ready to build.
Running software every week.
Not a status update. Something you can open and click, against your data as soon as access allows.
Bad news in week four, not month six.
If the premise looks wrong, you hear it early, in writing, with what we saw and what we tried. A well-argued no is a cheap outcome. A late maybe is not.
You own it and can run it without us.
Your repository, your infrastructure, your credentials. The handover call is short because nothing in it is a surprise.
Ishita Samadhiya
President
Aniruddh Mohan
President
Harrison Tang
Vice President
Sritej Bommaraju
Vice President
Tell us the narrow version of your question.
The most useful first message is one problem, one sentence about what a good answer would let you decide, and whatever data access is realistic. We will tell you whether it is a fit, and say so plainly if it is not.