There is a particular kind of dead time every software engineer knows. The code is written. The pull request is open. The next decision belongs to a row of little status checks. A test runs. A container builds. A cache misses. The engineer changes tabs, answers a message and tries to remember what mattered when the green check finally appears. Aditya "JP" Jayaprakash built a company around that interval.
Blacksmith, the San Francisco company he started with Aayush Shah and Aditya Maru in 2024, sells infrastructure for continuous integration, or CI. Its first offer was crisp: change one line in a GitHub Actions configuration and move the job to Blacksmith's hardware. The company said the result could run up to twice as fast at half the cost. The apparent simplicity hid the work underneath - high-single-core-performance processors, fast colocated caches, isolated microVMs and a control plane shaped around workloads that arrive in sharp, impatient bursts.
For Jayaprakash, the pitch was never only about faster machines. At Faire, the wholesale marketplace where he worked on search and advertising infrastructure, he saw how a slow pipeline could delay a hotfix. When an ads system earns money minute by minute, waiting to validate a change acquires a visible price. CI was the unglamorous passage between an engineer's intention and a company's confidence. He wanted to shorten it.
“We started by picking the least glamorous job in the stack: CI.”Aditya Jayaprakash, 2025
An optimizer before an operator
Long before he was selling compute, Jayaprakash studied the mathematics of compromise. He earned a computer science degree at the University of Waterloo in 2019, then completed a master's in computing science at the University of Alberta. His field was theoretical computer science, specifically approximation algorithms: methods for finding useful, provably close answers to problems that are too expensive to solve exactly at practical scale.
With professor Mohammad R. Salavatipour, he worked on capacitated vehicle routing across graph classes including bounded treewidth, bounded doubling dimension and highway dimension. A preliminary version appeared at the SODA algorithms conference in 2022; the research was published in ACM Transactions on Algorithms in 2023. The topic sounds distant from a CI cloud. The mental posture is adjacent. Both ask where the constraint lives, what must remain true and which trade can unlock the rest.
Define the bottleneck. Preserve the guarantee. Look for leverage in the structure of the workload.
At Faire, theory met a large production system. Jayaprakash worked across search, then became a founding member of the ads team. He led engineering for ad indexing and retrieval and helped the product grow from launch to several million dollars in annual recurring revenue. Search taught him about relevance and scale. Ads added commercial consequence. A system was no longer elegant because it merely worked; it had to work quickly enough for the business depending on it.
The dinner test
Jayaprakash, Shah and Maru had met years earlier at Waterloo. Before Blacksmith, they spent roughly nine months exploring ideas. They did research, lost conviction, recognized poor fits and moved on. Their own account of the period uses the word “meandered,” which is more revealing than the cleaned-up founder story in which every earlier choice points neatly toward the company.
The idea that held arrived over dinner. The friends were complaining about CI at their jobs, an ordinary engineer ritual. Their companies occupied different parts of the market, yet each had encountered slow pipelines, rising compute costs and the labor of managing infrastructure that developers did not want to think about. The overlap gave the complaint weight. In late November 2023, they started experimenting with ideas that would underpin Blacksmith.
A few weeks later, while scrolling Twitter, they saw a post from Y Combinator partner Dalton Caldwell. They already wanted to pursue Blacksmith full time, but felt early and uneasy about taking the leap. Applying to YC became a way to force the calendar. Two weeks later Jared Friedman offered them a place in the Winter 2024 batch. They accepted immediately.
Moving to San Francisco meant improvisation. The furnished apartments YC suggested were gone, so the team found an Airbnb in Bernal Heights, near a coffee shop they remembered fondly. Their three cats and one dog could not come, which created a second logistical project. Then the batch supplied its own clock: group office hours every two weeks, direct feedback from Friedman and former Lever founder Nate Smith, and the social pressure of arriving with measurable progress.
The team launched before it felt comfortable. Batch mates became early users. A developer with a large Twitter following found the product and posted about it, creating a wave of inbound attention. By fundraising time, Blacksmith had strangers running their CI on the platform. The founders took about 70 calls in two weeks. Angels often moved after a short conversation; institutional funds wanted a market category and a thesis. The original pitch did not map cleanly onto those frameworks. They adjusted. Once a lead offer arrived, Jayaprakash wrote, conversion with other firms jumped from nearly zero to above 80 percent.
“We applied to YC on a gut feeling.”Aditya Jayaprakash, on Blacksmith's W24 decision
The product behind the one line
CI has an awkward demand curve. A team can be quiet, then merge a set of changes and ask for hundreds of jobs at once. The jobs are short-lived; developers expect machines immediately; caches matter; a minute of queueing can erase a minute saved by a faster processor. General-purpose clouds were designed for many shapes of work. Blacksmith's wager was that this one shape justified a vertical stack.
The founders chose consumer gaming CPUs because compilation and testing often reward single-core performance. They placed fast storage close to the compute and used Firecracker microVMs to isolate jobs. They built warm dependency and Docker-layer caches, encrypted storage and controls for secure job access. Later they added observability: workflow timelines, test analytics, flaky-test detection and the ability to enter a running job over SSH. The unit of value moved from a rented runner toward an answer: why did this change take so long, or fail?
More code creates more checkpoints
Conceptual workload indexbased on Blacksmith's reported trend
Multitenancy made the economics possible. One customer's violent spike can be absorbed against another's quiet period. More independent workloads smooth aggregate utilization. About six months after launch, Jayaprakash found an essay by Amazon engineer Marc Brooker that articulated the math more cleanly than the founders' early simulations. The discovery was a small founder pleasure: seeing an intuition earned through operations reflected in a formal model.
The product found buyers quickly. In early 2025, Jayaprakash said Blacksmith had reached $1 million in annual recurring revenue with four people and more than 250 customers; until the week of that announcement, he had been the only person doing sales. Four months later, he reported $3 million ARR and more than 400,000 builds a day. Blacksmith announced a $3.5 million seed round led by GV in May 2025, then a $10 million Series A led by GV that September. At the Series A announcement, the company said more than 12,000 developers across over 800 companies had run a cumulative billion vCPU-minutes on its infrastructure.
When agents fill the queue
AI coding tools sharpened the company's argument. Code could be produced faster, but every generated change still needed to compile, pass tests and earn trust. In September 2025, Jayaprakash said Blacksmith's existing customers were running 60 percent more CI per developer each quarter. His framing was blunt: if a change takes an hour to test, an agent that writes code 100 times faster has simply moved the bottleneck.
On Megaport's Uplink podcast in June 2026, he extended the thought. Agents can create far more pull requests than a person, and each request fans out into a pipeline. The desired end state is quiet infrastructure: validation that scales with the code supply, catches failures, explains them and stays out of the developer's attention when everything works.
Quiet infrastructure still fails loudly. After a control-plane outage on July 21, 2026 blocked customer workflows for several hours, Blacksmith published a detailed postmortem. It described a strained Redis dependency, inadequate alerting and a burst of duplicate, out-of-order GitHub webhooks. The company listed changes to capacity, isolation and recovery. For a CEO selling confidence between code and production, accountability belongs inside the product story. The useful promise is the willingness to show the mechanism when the green check does not arrive.
The person inside the system
People who knew Jayaprakash before Blacksmith emphasize a combination that does not fit the lone infrastructure savant caricature. Rahul Pandey once organized ten engineers into a three-month side-project cohort. He later singled out Jayaprakash for shipping and monetizing his project because he understood the user. Pandey also noticed that Jayaprakash raised the level of the group by listening to other engineers' problems and offering relevant feedback. He eventually recommended the Blacksmith founders to YC.
That same range appears in Jayaprakash's public trail. He can write about approximation schemes, Dockerfiles, multitenant margins and fundraising psychology. He invited strangers interested in virtualization or bare metal to get coffee. He sold the first customers himself. His account of YC admits fear, awkward pitching and the leverage of peer pressure. The candor makes the operating philosophy easier to see: progress is a sequence of observable constraints, and embarrassment is rarely one of the important ones.
Blacksmith's ambition now reaches beyond faster GitHub Actions runners. The company describes a stack with compute at the base, observability in the middle and security at the top. Jayaprakash's stated mission is to help developers and agents merge code as quickly as possible. That goal is less theatrical than teaching an AI to write software. It may be just as consequential. Every new line still approaches the same gate, carrying the same request: prove that this change is safe to ship.
Follow the work
Jayaprakash publishes company and engineering notes through Blacksmith and shares operating updates on LinkedIn and X. His research and older code projects remain public as well.