WELCOME.txt

synaptech.sh — zsh

corrosion.vst.architecture

Bot platform architecture diagram
8
portfolio entries
ERP
MCP · bots · analytics
0
NDAs violated

HOW_I_WORK.md

01.
Specs before code.

Every engagement begins with a written technical spec. No spec, no code. This prevents scope drift and ensures we both know what "done" means before a line is written.

02.
Async-first communication.

I work in documented, asynchronous cycles. Weekly written updates, not daily standups. Your time is valuable; I respect it by being precise and self-directed.

03.
Honest scope assessment.

If your timeline is impossible, I'll tell you upfront. If your budget doesn't match scope, we'll renegotiate before starting. No surprises at invoice time.

04.
Ownership transfer by default.

You own everything built. Clean repos, documented APIs, runbooks included. I don't create dependency.

WHO_I_DONT_WORK_WITH.md

Spec-averse founders.

If "we'll figure it out as we go" is your engineering philosophy, we're not a match. Ambiguity is expensive and I refuse to bill you for confusion.

AI hype over AI utility.

I won't build LLM wrappers disguised as products. If you need "ChatGPT but for X" without a clear moat or defensibility thesis, I'll decline politely.

Equity-only, post-launch.

Pre-seed with a real prototype: possible. Series A company asking for "sweat equity": no. My rent is not a venture bet on your cap table.

Surveillance or dark patterns.

User tracking without consent, engagement manipulation, or deceptive UX. Hard no. I review use cases before signing any agreement.

click to expand
engineering philosophy
CORE_BELIEF.txt
"Software is not a product.
It is a decision
made permanent in code."

Every architectural choice, every abstraction, every shortcut — they compound. I treat code as a long-term liability, not a short-term deliverable.

authored by
Germán Martínez
Semi-Senior ML / Software Engineer
M.01

Complexity is a debt, not an asset.

Every layer of abstraction you add must justify its existence with measurable value. Teams collapse under architectures that impressed in a pitch deck but couldn't survive a Tuesday morning incident.

M.02

AI is infrastructure, not magic.

LLMs are probabilistic systems. They fail silently, drift over time, and hallucinate under pressure. Treating them like deterministic APIs is how you ship a liability. I build with evals, guardrails, and fallbacks.

M.03

The best system is the one that ships.

Perfection is the enemy of production. I optimize for correctness, observability, and maintainability — in that order. A working system with known tradeoffs beats an elegant system that never deploys.

M.04

Honest engineering over comfortable lies.

I will tell you when your idea is technically unsound. I will flag timeline risks before they become crises. I will document decisions I disagreed with. Candor is a professional obligation, not an option.

M.05

Ownership means leaving, not staying.

My goal is to make myself unnecessary. Every engagement ends with documentation, runbooks, and a team that can operate without me. Dependency is a consulting anti-pattern I actively refuse to create.

in practice, this means:
  • Written specs before any code is written
  • Eval suites shipped alongside every LLM feature
  • Architecture Decision Records for every major call
  • Honest scope and timeline pushback when needed
  • No vendor lock-in without documented justification
  • On-call runbooks before any system goes live
LLM / AI
OpenAI
Claude
Gemini
Llama / Kimi / Gemma
LangChain / LangGraph
Instructor / Outlines
DATA / INFRA
PostgreSQL / Redis
Pinecone / Weaviate
Kafka / Redpanda
Airbyte / Cognee
Feast (Feature Store)
ClickHouse
PLATFORM
Kubernetes / Docker
AWS / GCP / Azure
Railway / Fly.io
Terraform
GitHub Actions
Modal / Replicate
OBSERVABILITY
Langfuse / Helicone
RAGAS / DeepEval
Prometheus / Grafana
Datadog
Sentry
OpenTelemetry
P.01

Boring tech is a feature.

PostgreSQL before a distributed database. A cron job before a message queue. The most maintainable system is the one your next engineer can debug at 2am without Slack.

P.02

Evals before deployment.

Every LLM system ships with a benchmark suite. If you can't measure it, you can't improve it—and you definitely can't sell it to an enterprise customer with a legal team.

P.03

Runbooks, not tribal knowledge.

Every system I build ships with operational documentation. On-call runbooks, architecture decision records, deployment guides. Your team shouldn't need me to keep the lights on.