Luis Cruz — Backend & AI Engineering Portfolio

The hub of my engineering portfolio: two production-grade backend systems and a personal site. This repository holds the plan, the standards and the roadmap behind them — not the code.

Es: El hub de mi portafolio. Aquí viven el plan, los estándares y el roadmap; cada proyecto tendrá su propio repositorio cuando su primera tarea esté commiteada.

---

What this repo is (and what it is not)

This is not a project. It is the single source of truth for three projects that do not share code:

• the plan — what gets built, in what order, and why
• the standards — the rules that define "done" for every project
• the roadmap — the honest state of every epic and task

Each project gets its own repository, created when its first task lands. Nothing here claims to work yet. There are no screenshots of systems that do not exist, and no links to repositories that have not been created.

The three deliverables

Project	What it is	Core language	Target market
Caliber	Quality lab for search and RAG: load a corpus and golden questions, run comparable experiments (text vs vector vs hybrid, chunking, embeddings) and publish reproducible metrics — recall@k, MRR, nDCG — through an evidence console that shows everything in 60 seconds.	Python (FastAPI)	Spain
Continuum	Durable execution engine written in Rust: declarative workflows that survive crashes (event-sourced journal), retries, timers, signals and exactly-once effects verified by crash-injection tests — with published benchmarks and a chaos-button demo.	Rust (Axum/Tokio)	Switzerland / Germany
Personal site	Terminal-style interactive site presenting both systems with their real numbers, a downloadable CV (ES/EN/DE) and a technical blog.	TypeScript (Next.js)	Both

The two systems are independent. Each one starts with a single docker compose up and boots with pre-seeded demo data — never an empty screen.

The problem

I analysed 100 real backend job postings (LinkedIn, Spain + Switzerland/Germany) before writing a line of code. Three findings shaped the plan:

Finding	Evidence
The Spanish market is Python-first, and AI is no longer optional	Python appears in 26.5% of postings, Rust in 15, AI/LLM signals in 11 (10 of them in Barcelona).
The Swiss/German market is the opposite	Rust in 70.6% of postings, Python in 0. A Python-only portfolio does not open that market.
My GitHub could not prove the claims I was making	0 of my repositories had CI, topics or descriptions, while my profile advertised "CI/CD with GitHub Actions". Backend Rust appeared nowhere.

So the plan targets the real gaps: reproducible evaluation, durable systems with measured performance, CI that actually runs, and Rust that does real work — instead of another CRUD application.

How the pieces fit

                        ┌──────────────────────────────┐
                        │  Portafolio (this repo)      │
                        │  plan · standards · roadmap  │
                        └───────────────┬──────────────┘
                                        │
        ┌───────────────────────────────┼───────────────────────────────┐
        │                               │                               │
        ▼                               ▼                               ▼
┌───────────────────┐         ┌───────────────────┐         ┌───────────────────┐
│      Caliber      │         │     Continuum     │         │   Personal site   │
│  Python · FastAPI │         │   Rust · Tokio    │         │  Next.js · TS     │
│  pgvector + FTS   │         │  Event-sourced    │         │  Terminal UI      │
│  Eval console     │         │  Chaos-button demo│         │  Blog + CV        │
└───────────────────┘         └───────────────────┘         └───────────────────┘
      docker compose up             docker compose up          static build

No shared runtime, no shared database, no coupling: any project can be reviewed on its own.

Key decisions

Decision	Alternative rejected	Why
Build everything from scratch	Extend the previous repositories	Clean ownership: standards from the first commit and every line explainable in an interview. The previous repositories stay as history.
Evidence-first fronts with a 60-second guided tour	README-only presentation	Reviewers look at a project for about a minute: each system ships a front that proves its capabilities and a tour that triggers the key actions itself.
Caliber: reproducible experiments with a stable config hash	One-off demos	If a result cannot be reproduced, it is not published. Same config → same numbers.
Continuum: event-sourced journal as the source of truth	In-memory state plus snapshots	Crash recovery becomes a property of the design, not a patch. Exactly-once effects are verified by crash-injection tests in CI.
Measured numbers only	Absolute benchmarks without a baseline	p95, throughput, recall@k, recovery time — each with methodology and hardware. A number without a baseline is marketing.
English interfaces, Spanish learning docs	Trilingual product UIs (ES/EN/DE)	The market data does not ask for trilingual products; the learning documentation (in Spanish) is what makes the work defensible in an interview.
Zero budget	Paid cloud / LLM APIs	Local embeddings, free-tier models, free deployment. A reviewer can reproduce the whole stack.

Status

Honest state as of October 2026. Every epic is tracked in its own docs/PLAN.md.

Project	Epics	Tasks	Done	Next milestone
Caliber	8	29	0	Foundations + CI (Oct–Nov)
Continuum	8	26	0	Foundations + CI (Oct–Nov)
Personal site	8	34	0	Foundations + CI (Oct–Nov)

No measured results are published yet, because nothing is measured yet. The rule for every project in this portfolio is: if it is not measured, it is not published. Numbers — p95 latency, recall@k, throughput, recovery time, coverage — appear in each project README as its epics close.

Full roadmap — 89 tasks across 24 epics

Caliber — 8 epics · 29 tasks

#	Epic
0	Foundations — monorepo, dev compose, CI from the first commit, app factory + health
1	Corpus and golden questions — models + migration, deterministic seed (~120 docs, ~80 questions), chunking
2	Retrieval engine — full-text, pgvector, RRF hybrid, stable config hash
3	Experiments and metrics — recall@k / MRR / nDCG, async runner with SSE progress, optional LLM judge, leaderboard
4	API — API keys, corpus/experiments endpoints, live compare
5	Evidence front — live experiments, leaderboard charts, side-by-side comparator, 60-second guided tour
6	Observability and quality — logs, traces, Prometheus, coverage ≥80%, adversarial API tests
7	Deploy and close-out — multi-stage images, free VM deploy, README with real metrics, ADRs, learning docs

Continuum — 8 epics · 26 tasks

#	Epic
0	Foundations — workspace, CI (fmt/clippy/deny), dev compose
1	Model and journal — typed workflows, event-sourced journal in PostgreSQL, demo seed
2	Scheduler and execution — bounded queues with backpressure, retries, durable timers, signals, exactly-once effects
3	Durability and recovery — crash recovery, crash-injection tests in CI, deterministic replay
4	API and observability — Axum REST, live journal stream (WS/SSE), Prometheus
5	Evidence front — live mission control, chaos button, journal viewer, benchmarks panel, 60-second guided tour
6	Benchmarks — Criterion, reproducible methodology, results in the README
7	Deploy and close-out — lean images, guarded public demo, ADRs, learning docs

Personal site — 8 epics · 34 tasks

#	Epic
0	Foundations — Next.js scaffold, CI from the first commit, layout, performance baseline
1	Terminal engine — command parser, section registry, base commands, accessible navigation
2	Real content — single typed data layer, projects with TODO(metrics), integrity test
3	Sections — welcome, about, projects, skills, experience, contact, blog
4	Visual identity — sidebar, design tokens, mobile-first, chrome
5	Motion — typewriter, section reveals, performance discipline
6	Quality — accessibility, SEO, Lighthouse ≥95, real-device test
7	Deploy and close-out — Vercel, downloadable CV, learning docs, GIF and video

Engineering standard

These rules apply to all three projects. They are the definition of done, not suggestions.

1. CI from the first commit — lint, type check, tests and build on every push, with real services (PostgreSQL, Redis), not mocks.
2. Evidence-first README — the first screen answers in 10 seconds: GIF, live demo link, and a table of real metrics. Then pitch → problem → architecture → trade-offs → measured results → stack → how to run it → suggested review path → honest limitations → license.
3. ADRs in docs/decisions/ — context, decision, alternatives considered, consequences. Minimum four per project.
4. Measured metrics only. No invented figures, no skill bars, no percentage lists.
5. One-command demo with pre-seeded data, plus a 45–90 s video and a 10–15 s GIF.
6. Evidence front with a 60-second guided tour — every screen proves a capability; the tour triggers the key actions and narrates the technology behind each step.
7. Hygiene — no secrets in history, correct .gitignore, LICENSE, topics, single main branch.
8. Learning documentation — every epic ships a written explanation of what was built, why, how to verify it, and the five questions an interviewer would ask.

Suggested review path

This repository is short enough to read in five minutes. In this order:

1. The three deliverables (top) — what each system is and which market it targets.
2. The problem — the 100-posting analysis that decided the scope. Every claim carries a number.
3. Key decisions — seven trade-offs, each with the alternative that was rejected and the reason.
4. Status and roadmap — the real epic and task breakdown, including what is not built yet.
5. Engineering standard — the eight rules that define "done". This is what the projects get measured against.
6. Limitations — what this portfolio does not do. Read this one before judging the rest.
7. LICENSE — MIT.

When the project repositories exist, each README opens with a review path of its own: run the demo, read the critical file, run the tests, read the ADRs.

Limitations

• The systems are not built yet. This repository currently contains the plan, not the outcome. The roadmap and task counts above are the honest state.
• The market research is a snapshot of 100 postings collected in August 2026. It is directional, not statistically rigorous, and the raw dataset is kept private.
• Agent-assisted implementation. The code in the three projects is written with AI agents under my direction. The learning documentation exists precisely so I can explain and defend every design decision — a rule I hold to: if I cannot explain it, it does not get committed.
• Zero cost constrains the demo. Free-tier models and local embeddings mean the deployed demos are slower than a paid production setup. This is a deliberate trade-off, documented in each project.

Archive

Earlier planning was superseded twice and is documented rather than silently dropped:

• The LedgerCore / TickForge specifications (August 2026) and the NexusBoard / Cortex plans (September 2026) were superseded in October 2026 by Caliber and Continuum, after a market-fit review. They were retired from this repository; the reasoning lives in the Key decisions section above.

License

MIT — see LICENSE.