EPAM Systems is looking for a Data Solution Architect.
Let’s get straight to the point. EPAM is a behemoth. A global IT giant that drives digital transformation for other giants – from healthcare to retail. They have thousands of projects, tens of thousands of engineers, and all of it must operate like a well-oiled conveyor belt. Their business is selling expertise, packaged into teams and solutions.
So what pain are they trying to alleviate with this vacancy? Imagine this conveyor belt. Every day, a new client arrives at its input with a heap of problems: “We have data in a hundred systems, we want BI,” “We need an IoT platform with predictive maintenance,” “We want to monetize our data, but we don’t know how.” And each time, an experienced person is needed to untangle this mess, draw a beautiful diagram in Miro, select the tech stack, assess risks, and tell the team: “Dig from here until lunch.” This person is an architect. They are the bottleneck. Their time is expensive, their experience is unique, and finding such a specialist with 7+ years under their belt means months of interviews and a salary race. EPAM isn’t just looking for a person; they’re looking for a scalable brain for their solution-production conveyor belt.
Now, let’s imagine a dialogue between two directors at EPAM.
— Mikhail, we still can’t find a data architect for the retail project. The market is empty. Those who are available ask for so much that the project’s margin goes down the drain.
— Oleg, the same old story. What did you expect? It’s a bespoke item. Each of our architects is a walking knowledge base of our past successes and failures. You can’t just copy them.
— Exactly! You can’t copy them… But what if… you could? What if we created an “Architect-GPT” that knows everything about our past projects, all our technology radars, and best practices? So that a new project doesn’t start from a blank slate, but with a dialogue with a machine that can sketch out 80% of a ready solution in 5 minutes.
Fantasy? Not entirely. The task that a living person would solve for huge money can be delegated to an AI-based system. Not fully replaced, no. Rather, it would give the remaining architects superpowers, transforming them from “diagram drawers” into “idea validators.”
Here’s how it could look in practice.
Step 1. Creating a “corporate brain.”
The first and most important thing is to gather knowledge. EPAM has a gigantic legacy: internal Confluence, Jira, code repositories, architectural design records (ADRs), client presentations. All of this is fuel for AI. We take all these terabytes of unstructured data and run them through embedding models, creating a vector knowledge base. Essentially, we translate all the company’s experience into a language understandable to a neural network. This isn’t just keyword search; it’s semantic search. We add industry best practices, guides from AWS, Azure, GCP, and articles from Martin Fowler.
Step 2. Developing an “AI Solution Architect.”
This isn’t just a chatbot. It’s a system consisting of several components:
Interface: A project manager or business analyst describes the client’s business problem in a simple dialogue window: “We need a platform for analyzing customer behavior in e-commerce. Sources: CRM, Google Analytics, mobile app. Requirements: GDPR, real-time dashboards, predictive churn model. Budget – medium.”
Orchestrator (based on LangChain or a similar framework): It receives the request and “thinks” like an architect. First, it queries the vector database (RAG – Retrieval-Augmented Generation technology) and finds 3-4 similar projects EPAM has done in the past. From these, it extracts key architectural decisions, tech stack, and pitfalls.
Generative core (GPT-4, Claude 3 Opus): Based on the found information and the initial request, the core generates several artifacts:
1. Preliminary architectural diagram (e.g., in Mermaid or PlantUML format, which are immediately rendered into an image).
2. A list of recommended technologies with justification (“For streaming processing, we use Kafka, not Kinesis, because the client has expertise in self-hosted solutions”).
3. An estimate of effort and a preliminary calculation of cloud infrastructure costs (via AWS/Azure calculator APIs).
4. A list of potential risks (security, compliance, performance).
Step 3. Human-in-the-Loop.
And this is where the human enters the scene. But not the one sought in the $150k+ vacancy, but a senior architect remaining on staff. They don’t spend weeks gathering requirements and drawing first drafts. They receive a ready, 80% developed document from the AI. Their task is to review it, ask clarifying questions to the machine (“Why didn’t you suggest using Databricks? The client has a license”), make expert edits, and give the final “go-ahead.” Their productivity increases 5-10 fold. They can oversee not 2-3 projects, but 10-15.
How to ensure the AI isn’t suggesting nonsense?
Validation is key to trust. Here are a few ways:
1. Retro-testing: “Feed” the system descriptions of 20-30 successfully completed projects from two years ago. Compare the architecture proposed by the AI with what was implemented by humans. How closely do they match? Where did the AI make a mistake, and where might it have suggested a more optimal solution?
2. Parallel run: For a new project, give the task simultaneously to the AI and a team of architects. Compare the results after 3 days. Evaluate not only the quality of the solution but also the speed of its generation.
3. Quality metrics: Track how project metrics change where the “AI Architect” was used. Did the project start time decrease? Did the number of architectural errors found during development decrease? How accurate did the initial budget estimates become?
Instead of searching a overheated market for another expensive specialist who will become just another cog in the conveyor belt, one could invest in a system that makes the entire conveyor belt smarter and faster. EPAM is looking for a person to scale expertise. And that’s precisely what AI does best today. Ironic, isn’t it?
Источник: https://www.linkedin.com/jobs/view/4381281797/
