Claude Certified ArchitectCefalo
Present3 yrs 11 mosDhaka
Independent market intelligence for green certificates. Over 300 firms across the Nordics and Europe price and benchmark against its data.
- Carbon
- Guarantees of origin
- Power purchase agreements
- Renewable fuels
- 200+pipelines in production
- ~580×faster core lookup
- 12colleagues taught agentic AI
Present
Staff Software Engineer
- On Cefalo's AI pilot, taught 12 colleagues to use agentic AI tools like Claude properly — coding agents, agentic workflows, MCP tooling and prompting that holds up under real work.
- Sit on Cefalo's AI Task Force, shaping how the company runs its AI transformation.
- Built MCP tools that give agents first-class access to the systems behind Veyt's projects, rather than leaving them to guess.
- Designed agentic workflows that take the repetitive work off the team — the recurring tasks that quietly consumed hours every week now run themselves.
- Hardened a production auto-fix agent across ~50 automated repairs, closing the gaps in how it checked its own work and lifting the unattended fix rate towards 85%.
- Set the technical direction for data engineering work across the Veyt account.
- Built an HR agent for Cefalo at a company hackathon — an MCP-based RAG pipeline over the policy documents and the employee database, on LangChain, LangGraph, Pinecone, FastAPI and a local Ollama model.
Show all 7Show fewer
Senior Software Engineer · Data Engineer
- Architected and scaled the ingestion layer that pulls from dozens of external registries and APIs, and feeds the analyst dashboards, forecasting and decision tooling built on top of it.
- Built and run 200+ data pipelines producing the timeseries, timeseries groups and forecast data behind the platform's market intelligence.
- Cut processing time on the slowest pipelines by 50%, by moving the work to multiprocessing and removing the code paths that made it serial in the first place.
- Partitioned the large datasets so ETL runs in parallel, which brought down both the latency and the memory footprint of the high-volume executor jobs.
- Rewrote the queries and indexes behind the production timeseries API: a core fragment lookup went from 1,141 rows read to 19, and from 99.4 ms to 0.193 ms — about 580 times faster — taking database compute from ~1.75 to ~0.91 CU-h per hour.
- Put Redis in front of the heaviest reads and cut query response times by close to 60%.
- Reworked the Dockerfile and brought build times down by 40%, which took a slow feedback loop out of every deploy.
- Orchestrated every ETL workflow with Prefect, containerised the jobs, and shipped them to GCP Cloud Run through GitHub Actions.
