Proxet makes the complicated work.

A GIF is an absurdly small thing to bring down a system. A raised eyebrow. A shrug. Three seconds of somebody falling off a chair. Yet multiply those requests across enough conversations and the little search box becomes a formidable engineering problem. Giphy’s original PHP architecture had trouble keeping up with demand and finding the image a person actually meant.

Proxet helped rebuild the machinery. Its engineers worked on Scala microservices, search taxonomy, recommendations, and integrations. According to Proxet’s case study, the resulting system could handle 100 billion requests a month, ten times its earlier capacity. API timeouts fell by 95%. The joke arrived. The infrastructure disappeared from view.

A GIF is a tiny stress test

The revealing detail is where the experience came from. In the case study, former Giphy CTO Anthony Johnson points to Proxet’s background in advertising technology, another business in which huge numbers of requests must move quickly. Expertise acquired in one market became useful in another. A GIF search and an advertising system share a clock.

That helps explain the company better than a catalogue of programming languages. Founder and CEO Vlad Medvedovsky’s firm sells the ability to recognize an engineering pattern, assemble the people who understand it, and stay involved through implementation. Its September 2026 announcement describes more than 350 technologists and over 200 client engagements.

The hospital had a dependency problem

At Mass General Brigham, the friction looked different. Its Insight Research Portal supported research administration, with modules for agreements, human research, biosafety, and other responsibilities. Each module had its own rules. Some depended on others. Improving one piece could become an expedition through the rest.

Proxet’s account describes modernization efforts stalled for more than a decade, brittle architecture, and long release cycles. The engagement began with interface work. As new requirements emerged, the remit expanded into backend redevelopment, rewritten APIs, and a gradual move toward microservices and Azure. Proxet also introduced continuous integration, delivery practices, and automated testing.

The sequence matters. A new interface exposed the need for deeper changes; the work followed those dependencies. The company reports that pages previously taking more than twenty seconds loaded in one or two seconds afterward. Meanwhile, moving off physical servers released resources previously tied to maintenance. Better architecture and a better way of releasing software reinforced each other.

The spreadsheet bill comes due

General Catalyst supplies a less theatrical example. Its portfolio reports covered more than 400 companies. Producing them required over 100 person-hours, according to Proxet, with Excel helping consolidate scattered data. The expense was embedded in recurring effort: assembling, reconciling, and checking information before anyone could use it.

Proxet built a Google Cloud platform that collected information from systems including Salesforce, Aumni, AllVue, and Affinity. Airbyte moved it, BigQuery stored it, dbt transformed it, and Power BI presented it. A custom management tool let users resolve conflicts. The reported gains included less manual work and fewer errors in valuations and tax accounting.

Readers can copy the logic without copying the stack. Identify the report people keep rebuilding. Trace each number to its origin. Give disagreements a place to be resolved. Automate the route only after the definitions make sense. A faster pipeline carrying inconsistent figures simply delivers the argument sooner.

At Zus Health, access to real patient records created a different obstacle: privacy requirements limited the data available for development. Proxet’s solution included a synthetic test-data generator, mapping between inconsistent sources, and a patient-matching algorithm using two identifiers. Its case study reports onboarding 47 digital healthcare companies. Here, the useful engineering intervention was making it possible to test the product while protecting the information it handled.

A platform still needs a builder

This is Proxet’s position in the market: a services firm working between technology platforms and the organizations trying to use them. Customers buy discovery, architecture, implementation, dedicated teams, and support. Its managed teams cover testing, documentation, and maintenance alongside development. The commercial relationship can continue after launch.

The range includes Azure modernization, Databricks pipelines, and Palantir Foundry implementation. Proxet announced its Palantir partnership in May 2024. Its Aurora Payments work produced ARISE, a proprietary payment platform with custom APIs and integrations. These are customer systems built for particular needs. The appeal is strongest when an organization needs specialized engineering capacity across several connected problems.

Code got cheaper. Clarity did not.

The latest offering carries that argument into AI. In September 2026, Proxet launched its Intent-Driven Lifecycle service. The proposal combines shared organizational context, AI-assisted execution, outcome verification, and changes to delivery roles. An executive workshop identifies bottlenecks before a bounded delivery stream tests the economics.

Its Metadata project makes the distinction visible. The advertising platform already had sophisticated campaign tools, but customers faced a steep learning curve. Proxet helped connect separate agents and actions into a Claude-powered conversational workflow. Human review remained part of campaign launches and other consequential actions. The challenge had moved from giving an assistant individual capabilities to making those capabilities cooperate reliably.

The cost question belongs at that level too. Proxet’s Claude directory spans engagement budgets from under $25,000 to more than $2 million. Those are broad scope bands. A buyer needs a baseline, a bounded assignment, and an agreed result before that range becomes meaningful.

“When code generation becomes trivial, the real bottleneck in software delivery becomes human clarity and validation.”

Vlad Medvedovsky · September 2026

The promise depends on those last two words. Teams still need agreement about the problem, permission to access the relevant data, and a credible way to judge the result. For a simple need already met by standard software, bespoke engineering may add needless work. For tangled systems, the test is more concrete: can people finish the task with less friction? At Giphy, that meant getting a joke through.