Vlad Bondarenko · AI Engineer

Background

6+ years shipping production ML and data systems: LLM applications, fine-tuned translation models, large-scale data pipelines and time-series forecasting. BuiltAI (Lead AI Engineer), CarbonChain (Customer Engineer → Senior Data Engineer), THG (Data Scientist), Sheffield Solar (solar forecasting integrated into National Grid). Deep Python, GCP and AWS.

Skills

AI engineering

LLM application engineering4 years
Evals & reliability3 years
Retrieval / RAG3 years
Agents & tool use2 years

Software

Python9 years
TypeScript / Node5 years
Distributed systems6 years
Cloud & infra (GCP + AWS)6 years

Projects

A couple of representative engagements. Outcomes first; no logo wall.

2025 –

LLM customer onboarding & ingestion

Ongoing customer uploads went from ~10 hours to under 10 minutes.

Context: BuiltAI, Lead AI Engineer. Customer data arrived as arbitrary spreadsheets: no fixed schema, columns that drifted between uploads, and a manual reconciliation step every time.
Approach: Built automated drift adjustment so the pipeline adapts when a customer changes their layout, codified the parsing rules so they are versioned and testable, and put a fast QA loop around the whole thing so a bad parse is caught before it lands.
Outcome: Ongoing uploads: ~10 hours → under 10 minutes. The team spends its time on exceptions rather than on every file.

2025 –

AI-assisted development across the eng team

The team went from roughly weekly releases to shipping multiple times a day.

Context: BuiltAI. Engineers were using coding assistants individually, with no shared way to keep generated code safe to merge or to keep non-technical colleagues in the loop on what was being built.
Approach: Standardised a setup where specs and shared rules are the source of truth for LLM-generated code, enforced by security and reliability gatekeepers (type checking), a TDD-style workflow and a streamlined CI pipeline. Specs doubled as the bridge to non-technical colleagues. Ran company-wide workshops so the practice stuck.
Outcome: Releases went from ~weekly to multiple times a day. Related: the core calculation engine was re-architected as a DAG (large portfolios ~1 minute → seconds) and, because its behaviour was captured in specs, rewritten in Rust in weeks against an original ~6-month estimate.

2022 – 2025

Supply-chain ingestion at scale

Customer onboarding went from ~30 days to ~12; ingestion from weeks to hours.

Context: CarbonChain, Senior Data Engineer. Commodity traders supplied their supply-chain data as messy spreadsheets, CSVs and PDFs, each in its own shape, with counterparties named inconsistently across sources.
Approach: Built a generic pipeline that turns those inputs into a standardised supply-chain graph, with an entity-resolution layer over the noisy counterparties so the same company resolves to one node regardless of how it was written.
Outcome: Onboarding ~30 days → ~12. Ingestion weeks → hours. Re-ingestion 7 days → 16 hours.