Agent control
Durable scope, sandboxed workspaces, explicit status, and review gates keep long-running agent work bounded and inspectable.
Hello, I’m
Saurabh Shubham
AI Platform · Data Engineering · Berlin
I build controlled agentic workflows on production data foundations.
My work combines bounded execution, deterministic acceptance checks, human review, and recoverable delivery with 7+ years of Python and SQL engineering across pipelines, orchestration, lakehouse systems, and cloud platforms.
Experience 7+ years in production engineering
Current Data Engineer at GROPYUS
Live build Regulation Check
01
What I bring to AI infrastructure
Data foundations
Python, SQL, CDC, orchestration, and lakehouse patterns provide the reliable state and data movement AI workflows depend on.
Recoverable delivery
Tested releases, health checks, startup retry, pinned artifacts, and rollback make deployment failure visible and recoverable.
Selected engineering evidence
Proof, not pitch
One live product, one workflow architecture, and one recent learning project—each labelled by what is implemented.
01 / Live product
Regulation Check
A live AI governance workspace where an LLM turns free-text system descriptions into reviewable structured facts, while deterministic screening and sourced evidence keep decisions traceable.
Open Regulation Check- LLM intake
- Vercel AI Gateway extraction into a strict AI-system inventory schema
- Guardrails
- Untrusted-input boundaries, schema validation, provider fallback, and human confirmation
- Platform
- FastAPI with tenant-isolated Supabase PostgreSQL, evidence, tasks, and reports
- Agent tooling
- MCP-connected tools for bounded research, repository access, and delivery verification
- Decision path
- Deterministic screening with sourced obligations
- Operations
- Approval-gated CI/CD, pinned Docker images, health checks, startup retry, and last-known-good rollback
02 / Architecture design
Pasin
An independent multi-provider orchestration design for long-running agent work. Presented as architecture—not a claimed live product.
- Scope
- Durable task definition before execution begins
- Isolation
- Sandboxed workspaces for bounded agent changes
- Verification
- Deterministic acceptance checks before human review
- Recovery
- Git milestones designed for inspection and rollback
03 / Recent learning project
Retail Demand MLOps Demo
A hands-on model lifecycle project connecting production data-platform experience with ML-specific delivery and operations.
View the GitHub project- Training
- Time-based split, schema checks, scikit-learn pipeline, and reproducible synthetic data
- Lifecycle
- MLflow tracking, model versions, candidate and champion aliases, and a quality gate
- Serving
- FastAPI inference, readiness, model-version reporting, and Prometheus metrics
- Operations
- Drift checks, Docker, CI, runbook, architecture decision record, and incident review
02
Production experience
-
GROPYUS
Data Engineer · Berlin
Build and optimize Python data pipelines connecting robotic manufacturing systems with enterprise platforms across CDC, lakehouse, and graph-database architectures. Model and orchestrate production workflows with Dagster, Airflow, dbt, and DLT; ship tested changes through Azure CI/CD.
-
Sigmoid
Software Development Engineer · Bengaluru
Built repeatable ETL pipelines for commercial sales reporting with Python, PySpark, Pandas, and Airflow. Provisioned delivery infrastructure with Terraform across Google Cloud and AWS.
-
Amdocs
Software Engineer · Pune
Translated CRM requirements and wireframes into Java and Spring backend capabilities. Developed REST and SOAP service interfaces from implementation through integration.
03
Technical stack
Agent systems
Languages
Data systems
Platform + cloud
Engineering judgment
Good systems make failure boring.
I enjoy turning ambiguous work into explicit state, testable boundaries, and an operating path another engineer can understand.
The happy path is only the start. I care about partial data, stale state, failed startup, review handoffs, and the route back to a known-good release.
Questions before I call work done
- Can another engineer trace the state?
- Is failure visible, bounded, and recoverable?
- Does verification test the promised outcome?
04
Education
Birla Institute of Technology, Mesra
B.E. Computer Science
AI platform or data engineering role?
I build the control and data layers behind reliable AI work.
Agent orchestration, AI governance, production pipelines, and the infrastructure connecting them.