04 / about
Victor
I'm a Data & AI engineer who works end-to-end: ingestion and sync engines, data reliability, multi-source architecture, auth and security, dashboards people make decisions with — and, on my own time, agents, local LLMs and quant research. Most of my work happens in direct-response e-commerce: gateways, affiliates, call centers, refunds and the money in between.
Correctness over vibes
A number that can't be reconciled doesn't ship, however good the chart looks.
Measure before claiming
Query the source first. Twice the “missing” integration was already in the database.
File-first, then DB
Versionable artifacts on disk; the database is a queryable mirror, not the source of truth.
Honest status and credit
In production, Evolving, Shipped, Partially remediated — and when work was shared, I say which part was mine.
Electrical engineering
Power-system studies: short circuit, protection and selectivity, electromagnetic fields — and the Python scripts to process the results.
Technical analyst
Solar-energy projects and Python tools that automated pre-sales: generation simulation, financial return, web automation.
Data analyst
Direct-response e-commerce: KPIs, dashboards and finding where sources and platforms disagree.
DataOps / data engineer
Pipelines across marketing, payment and finance platforms, a dashboard hub, data reliability — and agentic workflows for data work.
Data Engineering
- Postgres & SQL
- Sync engines & checksums
- Data quality invariants
- Legacy MySQL forensics
- Python ingestion & APIs
- AWS lake (Glue, Athena, S3)
Product & Frontend
- Analytics dashboards
- React + TanStack
- Tailwind / design tokens
- Edge deploys
Agents & LLMs
- Context engineering & MCP
- LangGraph multi-agent systems
- llama.cpp / vLLM serving
Quant & Research
- Portfolio optimization
- Regimes & event studies
- Monte Carlo & covariance
Auth & Security
- OAuth & roles
- Postgres grants & RLS
- Secrets management
01
Plan before code
Spec, contract and checkpoints first. Code is the cheap part.
02
Invariants, not hope
Checksums and cross-layer tests that fail loudly.
03
Durable state > timers
If a restart can lose it, it isn't state.
04
Small models where they fit
A quantized 9B on a laptop GPU covers a lot of agent work.
05
Humans approve, agents execute
Agents work inside guardrails; people own the decision.
Spreadsheet → Python quant engine
Low-code → durable services
Static reports → live dashboards
API LLMs → self-hosted models
Copilot → agent harness
→ let's talk