The Benchmark and the Bootstrapper
The QuestBridge letter arrived at a house where the lights might get cut off. Cody Coleman was born in prison - his mother incarcerated, his father already gone. His maternal grandparents took him in, two people living on Social Security checks trying to keep a kid pointed somewhere other than down. A high school teacher, watching him work, set up a private fund just to cover his lunches so he could stay in school.
Then a scholarship program found him. QuestBridge placed him at MIT, and Coleman did what people who have never had a safety net tend to do: he ran. He graduated with a BS in Electrical Engineering and Computer Science, became president of the EECS honor society, and left Cambridge knowing that the most interesting problems in computing were still unsolved.
He moved to Stanford for a PhD, landing in the lab of Matei Zaharia - the co-founder of Databricks and one of the architects of Apache Spark. The mentorship was formative. Zaharia thinks in data infrastructure. Coleman absorbed that framing and aimed it at the emerging messiness of machine learning systems.
"I want to show that anyone can be successful regardless of their background."- Cody Coleman, CEO, Coactive AI
The PhD work was not about building the next flashy model. It was about asking: how do we know when a model is actually fast? How do we compare systems without letting vendors grade their own homework? The answer was DAWNBench in 2018, a time-to-accuracy benchmark that gave researchers a shared measuring stick. Then, in 2019, came MLPerf - the version the industry actually standardized on.
MLPerf did not just become popular. It became the default. Google uses it. Meta uses it. Microsoft uses it. NVIDIA ships hardware results benchmarked against it. When a chip company says their accelerator is fastest for training a transformer, the number they're citing almost certainly came from a framework Coleman helped write as a PhD student in Palo Alto.
MLPerf - which Cody Coleman co-created during his Stanford PhD - is now the industry-standard benchmark for machine learning hardware and software performance. Every major AI chipmaker (NVIDIA, AMD, Intel, Google TPU) publishes results against it. Coleman co-founded MLCommons, the nonprofit that stewards MLPerf's development.
What Coactive AI Actually Does
When Coleman finished his PhD in 2021, he co-founded Coactive AI with William Gaviria Rojas. The premise is simple to state and hard to solve: most enterprise data is visual. Images. Video. Product photos. Security footage. Sports broadcasts. Marketing assets. And virtually none of it is searchable, because search requires structure, and visual content has almost none by default.
Coactive's platform ingests those image and video libraries and makes them queryable - in plain English, without manual tagging or metadata preparation. Ask it to find "outdoor shots with warm lighting featuring a product in the foreground" and it finds them. The underlying technology stacks multimodal AI models trained on visual semantics, fed through a data pipeline architecture that Coleman knows well from his Stanford years.
The use cases break across media companies (searching broadcast archives), retail (product discovery and recommendation), sports analytics (play classification, highlight detection), and marketing operations (brand compliance, asset management at scale). Every one of those verticals has been drowning in visual data that their existing text-based search tools cannot touch.
"We're bringing structure to unstructured visual data so enterprises can unlock insights they've never had access to before."- Cody Coleman
In May 2024, Coactive closed a $30 million Series B, co-led by Emerson Collective and Cherryrock Capital, with participation from Bessemer Venture Partners, Greycroft, and Andreessen Horowitz. The round valued the company at roughly $200 million. Total funding stands at $44 million. The company is 63 people, headquartered at 60 South Market Street in San Jose, and operates in the space where enterprise AI meets the pile of visual content every large company has but cannot use.
In April 2024, Coactive launched MediaPerf - the first open benchmark for AI video understanding in the media industry. The instinct to measure before claiming is very much still there.
The Benchmarks He Built
The Data-Centric Thread
Coleman's research trajectory at Stanford had a consistent through-line: data matters more than people think. His 2019 paper "Selection Via Proxy: Efficient Data Selection For Deep Learning" argued that you don't need to label everything - smart selection outperforms brute-force annotation. His 2022 AAAI paper on "Similarity Search for Efficient Active Learning and Search of Rare Concepts" pushed that logic further.
The commercial insight at Coactive AI is essentially the same idea scaled up. Enterprises don't need to manually tag 10 million images. They need a system that understands visual content well enough to answer questions about it without the tagging bottleneck. The academic work and the product aren't separate tracks - they're the same conviction expressed in two different registers.
He gave a TEDx talk at Stanford titled "Digging Deeper: How a Few Extra Moments Can Change Lives" - a meditation on the human side of that same principle. The small decision to look a little closer, to invest a bit more attention, is what separates compounding from stagnation. It's also a decent description of how he built his career: granularly, carefully, with better-than-average instrumentation at each step.