Data Engineer (official title; current specialization: Azure) at IBM, on the Heineken global retail & beverages account since June 2022 — with applied AI systems engineering, built on IBM's Consulting Advantage (ICA) platform, as a differentiating second strength.

Selected outcomes across ~4 years on the account:

• An 80% cut in monitoring-dedicated headcount allocation (1.0 -> 0.2 FTE) across 200+ daily pipelines, from a self-designed Python-based alerting system that eliminated manual tracker-sheet monitoring.
• ~€1,000/month in storage cost savings from a solo-built Z-Ordering/VACUUM optimization initiative, and — as a separate, unlinked, billing-confirmed outcome — a >50% reduction in Databricks compute cost from a job-cluster right-sizing/idle-termination change.
• 50-75 engineering hours/month recaptured, combined, by two internal AI assistants I solo-built (DRS Operations Helper, DRS User Story Generator) — recognized by the client in a town hall — and, separately, 20-30 engineering hours/month recaptured by a third assistant ("Databricks Transform Expert") I later built to accelerate a gold-layer migration.
• That migration itself: 112 legacy notebooks consolidated into 6 SQL-based Delta Live Tables pipelines, delivered solo in 240 hours against a 324-hour internal estimate — cutting the pipeline count from 12 (each running 1-2 hours) to 6 (each now running 45-60 minutes).
• A Microsoft Graph API ingestion consolidation that cut runtime from 6-8 hours to 3 hours, with zero high-priority incidents in the 3 months since deployment (down from 3-4/month).
• A ~50% reduction in Azure platform-monitoring effort (40 -> 20 hours/month) after taking over ownership of the DataOps Observability Platform roughly 6 months into its build.

Outside IBM, I design custom multi-agent frameworks in LangGraph as personal projects (see below) — working through architectural trade-offs around context length, cost boundaries, state machines, and data security, rather than wrapping a single API call.

Jun 2022   ->   Associate System Engineer (graduate hire); ~2.5-month training
                v  program, then allocated to the Heineken account (Aug 2022)
Jun 2023   ->   Promoted to Associate Data Engineer (Heineken account)
                v  DataOps ownership, optimization initiatives, two internal
                   AI assistants built on IBM Consulting Advantage
Jan 2025   ->   Promoted to Data Engineer (Heineken account); specialization: Azure
                v  Platform reliability, gold-layer migration, a third internal
                   AI assistant, and DataOps Observability Platform ownership

Personal engineering work, built independently outside of IBM employment - not professional/job experience.

 LangGraph  Python  sqlite-vec  Docker  discord.py

A config-driven fleet of LangGraph agents in a single Discord bot, exposing 5 personal-assistant agents (Goal, Learning, Health, Finance, Research). (Single-operator system.)

• YAML config-driven engine - every agent is an instance of one shared ReActAgent engine, described declaratively in a YAML file, backed by per-user SQLite (relational) and sqlite-vec (semantic vector) storage.
• Agent loop - Decide -> Plan -> Execute <-> Tools -> Evaluate -> Final, a LangGraph state machine, with tool-call budgets and evaluation-loop caps.
• Key architecture - auto-generated typed CRUD tools per table (search/get/list/insert/update/delete) parsed from each agent's schema at startup; multi-tenant isolation (separate SQLite files per user, WAL mode, async per-user write locks); fail-fast Pydantic config; TTL+LRU agent caching; SQLite authorizer blocking DDL; prompt-injection redaction.
• Tech - Python 3.12+, LangGraph, discord.py, sqlite-vec, uv, Docker/Docker Compose.
• Testing - 248 tests, ~91% coverage, fully offline (mocks requests/OpenAI/ChatOpenAI/TavilyClient); gated by ruff lint/format.

 LangGraph  Python  FastAPI  Pydantic  tectonic

A multi-agent resume generator built on LangGraph. Takes a candidate profile and a job description, and a "council" of four persona agents (CEO, Hiring-Manager, Recruiter, ATS Specialist) plans, writes, and critiques a resume until it clears their score gates. (Single-operator system.)

• Workflow - three phases: parallel extraction (analyze_jd + parse_profile -> gap_analysis), parallel four-persona council planning joined by a conflict-resolution step into a final plan, and a write-and-refine loop (initial_write -> evaluator nodes -> edit_resume) capped at 3 cycles.
• Score gates - CEO >= 80, Hiring-Manager >= 80, Recruiter >= 85 (defined in config.py).
• Output formats - Markdown (zero-loss), LaTeX (LLM conversion), PDF (via the tectonic engine).
• Tech - Python 3.12+, LangGraph, Pydantic structured outputs (7 schemas), OpenRouter for LLM access, FastAPI web UI with SSE live progress.
• Testing - 51 passing tests, fully network-free (LLM stubbed via FakeLLM).

• Client-facing communication - presents architecture proposals and analysis directly to the client;
• Self-directed upskilling - independently learned Power BI to deliver a dashboard in the absence of a dedicated BI resource on the team; self-taught prompt engineering and applied AI/RAG fundamentals while building the AI assistants.
• Architecture governance - solo-designed solutions (retry orchestrator, cluster right-sizing, currency-ingestion pipeline, Graph API consolidation, Collibra migration, observability-platform work) consistently reviewed and approved by architects/team lead before production.

Languages

Data Platform & Engineering

Applied GenAI & AI Infrastructure

Microsoft Fabric & Analytics