Prod.Andres Ortiz — Systems reel
DirectorA. Ortiz
Roll01
Scene00
Take26
CameraA
FPS60
Date----.--.--
StatusMounting statue.glb — 87,547 trisBest with sound — switch it on, bottom right
000

Andres Ortiz — Software Engineer. AI systems · software & cloud architecture.

Systems, not demos.

Agents, MCP servers and gateways, tool governance, context budgets, evaluation that can fail — designed as systems, not demos.

Word 01/03
The recordEvery number here was taken from a real repository.
Fact 01/052026-09-25

363 unit tests guarding a world that is generated, not authored

363 — unit tests guarding a world that is generated, not authored.

SourceNOCLIP · tests/unit
Fig. A547 × 678
The engineer, under red light
04Profile— Who is this?

Ibuildsystemsattheintersectionofsoftwarearchitecture,cloudinfrastructureandartificialintelligence—andinvestigatewhatcomesnext.

Role
Software Engineer
Focus
AI systems · software & cloud architecture
Base
Quito, Ecuador — UTC−5
Languages
Spanish · English
GitHub
@AndresO7 ↗
  • 01 · CertifiedAWS Solutions ArchitectAmazon Web Services
  • 02 · CertifiedClaude CodeAgentic development
  • 03 · ProgramGitHub Campus ExpertGitHub Education · 2023
Fig. 01Knotted cords — one per contract
Selected work01/06

Quipu

Enterprise AI asset gateway

2026Own product · in progress

Learned

Put the invariant in one place.

RustTokio · AxumPython · FastAPINext.jsPostgres + pgvectorRepository ↗

Also built

  • 01KODAMulti-agent orchestration for DevOps workflowsLangGraph · A2A · MCP
  • 02Cloud BotNatural-language automation of AWS deploymentsGo · Kafka
  • 03Pico y PlacaQuito’s driving-restriction rules as a tested libraryTypeScript · Jest
  • 04PetShopTechnical test on the PetStore APINext.js 15 · React 19 · Zustand
  • 05Landing pagesIncluding ISCD, a security companyNext.js
06Method— How do you think?
METHOD

How the thinking runs — shown by quoting the documents it produced. Spanish originals stay next to the translation.

  1. 01

    Why does it break?

    Name the real constraint before choosing a stack.

    Quipu didn’t start with Rust. It started with a diagnosis: harnesses run out of context in four distinct ways. Each got its own defence, and the architecture followed from the defences.

    “Governance is context optimization.”

    La gobernanza ES la optimización de contexto.

    $ visible_tools = policy(role, workspace, assets)Quipu · docs/architecture.md
  2. 02

    How do the parts meet?

    Fix the contracts before the code.

    Nine contracts define how the gateway, optimizer, control plane and agents talk — before most of them exist. The seams are agreed first, so the parts can be built in parallel, by people or by agents.

    “Contracts are stable: changing one is a breaking change.”

    Contratos (estables — cambiarlos es un breaking change).

    Quipu · MASTER-PLAN §6.2
  3. 03

    What is already decided?

    Write decisions down, with the reason and the owner.

    Some choices are closed on purpose. Writing them down stops them from being re-litigated in every session — including sessions with AI agents that weren’t there when they were made.

    “Fixed decisions — don’t re-litigate without the owner.”

    Decisiones fijadas (no re-litigar sin el usuario).

    Quipu · MASTER-PLAN §4
  4. 04

    Where does it fail?

    Make the test able to fail.

    The end-to-end harness ships with a negative control: run it with the wrong allow-list and it has to go red. If it stays green, the harness is what’s broken.

    “A harness you have only ever seen pass has proven nothing.”

    Un harness que solo se ha visto pasar no ha demostrado nada.

    $ CURATED=add,get_time,echo bash dev/smoke-e2e.sh # must FAILQuipu · dev/README.md
  5. 05

    What is actually proven?

    State the honest limit.

    Deployment artifacts that were never executed are marked as verified by inspection only. Targets a phase can’t meet by construction are exempted — and the plan names the mechanism that will.

    “Verified only by inspection: they do not count as tested.”

    Verificados solo por inspección: no cuentan como probados.

    Quipu · MASTER-PLAN §13
07Context

Context is the resource

Task“Why did last night’s deploy fail? Open an issue with the root cause.”

Direct — 12 MCP servers, 96 tools, 6 skills0tok
200K
Through Quipu — the surface this identity may see0tok
200K
  1. 01 Assemble the context
  2. 02 Reason
  3. 03 Call a tool
  4. 04 Observe
  5. 05 Reason again
  6. 06 Observe
  7. 07 Ask before writing
  8. 08 Answer
Overflow — the harness compacts

96 tools with full schemas + 6 skills loaded in full — vs the 9 tools this role may use, as short signatures.

08Lab— Do you investigate?
Experiments
  • Question

    Can a procedurally generated building feel wrong — instead of just random, or just decayed?

    Hypothesis

    Wrongness is semantic. Noise reads as decay; a correct memory of a place with a few precise errors reads as dread.

    Result

    Twelve memory errors — a clock without hands, an EXIT sign on a blank wall, carpet climbing the wall — at most three per chunk. A reachability test guarantees no error ever blocks the route.

  • Question

    Does a procedurally generated reverb actually sound like the room it claims to be?

    Hypothesis

    If each room’s impulse response is generated per material, its RT60 per band should measure within spec.

    Result

    “Dark” materials measured bright. The tail is now three independent noise bands through Linkwitz–Riley filters, and per-band RT60 measures within a few percent of spec.

  • Question

    Should the Phase 0 walking skeleton already meet the data plane’s latency SLOs?

    Hypothesis

    A skeleton should be held to the final targets from day one, so regressions are caught early.

    Result

    A p50 ≤ 1 ms budget containing a network call is impossible arithmetic. Phase 0 is exempt from two SLO rows, and the plan names the snapshot that restores them.

  • Question

    How should a client’s 3D house configuration be saved?

    Hypothesis

    The obvious design: export the model — geometry, textures — and store the file.

    Result

    All geometry is procedural. A saved model is ~300 bytes of JSON, validated on read; the price is recomputed, never trusted from storage.

  • Question

    Is a deterministic, identity-projected tool surface better than letting the agent search for tools?

    Hypothesis

    Yes: a surface computed from identity costs fewer tokens and fails less than a search the agent has to drive.

    Result

    Pending — the Context Ledger (Phase 3) will count tokens saved per day, team and asset.

Archive

Earlier investigations, as the earlier portfolio records them.

  • 2024Generative AI: a new frontier in DevOps automationPublished research · ESPE
  • 2023Generating code from visual interfaces with LLMsAccepted paper
  • —CAG Engine — cache-augmented generation to spend fewer tokensResearch prototype
LineageFrom the model writing code to agents as a workforce.
01 / 052023

The model writes code

FunCodeGenerator: sketch a wireframe, and a vision model returns a working prototype you can annotate and iterate.

02 / 052024

The model operates infrastructure

Quasar: a local Llama 3 8B writes Terraform for AWS, treats apply errors as observations and loops until it works.

03 / 05THEN

Agents coordinate

KODA: multi-agent DevOps workflows with LangGraph, A2A and MCP — and research on cache-augmented generation.

04 / 052026

Context becomes the resource

Quipu: the gateway that decides what a harness gets to see — governance and context optimization as one mechanism.

05 / 052026

Agents become a workforce

NOCLIP: a game built by parallel agent workstreams under contracts, verifiers and independent research passes.

10Harness— How do you work with AI tools?

Agents need
a harness

 0                   1                   2                   3
 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                           OBJECTIVE                           |
+---------------------------------------------------------------+
|                    FILES TO CREATE / TOUCH                    |
+-------------------------------+-------------------------------+
|      CONTRACTS CONSUMED       |      CONTRACTS PRODUCED       |
+-------------------------------+-------------------------------+
|    TDD: test first · watch it fail · minimal code · green     |
+-------------------------------+-------------------------------+
|     VERIFICATION COMMAND      |      ACCEPTANCE CRITERIA      |
+-------------------------------+-------------------------------+
Task packet — drawn the way RFCs draw packet headers

NOCLIP was built by five parallel workstreams — level generation, the camcorder post pipeline, audio, entities and UI — each against its own dev harness. Verifier passes flagged: log soft-lock · seed links · Lessee ambient · pacing · line of sight.

  1. 01

    Spec before code

    Brainstorm → an approved spec → a phase plan of bite-sized TDD tasks. The plan is the entry point for any new session, human or agent.

  2. 02

    One fresh agent per task

    Each task goes to a new subagent with only its packet, then a two-stage review before the next one starts.

  3. 03

    Verification before completion

    “Done” means the commands ran and the evidence is shown — tests, lint, the deliverable actually running. Never a claim.

  4. 04

    Agents need eyes

    Playwright with fixed seeds and scripted routes: screenshots, FPS logs and a debug hook let agents verify what they built.

  5. 05

    Two agents, one question

    For open design problems, independent research passes from different agents against the same build — then a synthesis into contracts.

  6. 06

    Memory is a file

    Lessons go into NOTES.md — one per entry, updated instead of duplicated — and CLAUDE.md / AGENTS.md live in every repository.

11Handshake— What could you do for us?

Have a system worth investigating?

I’m looking for teams building AI systems that have to work in production — where architecture, cloud and AI can’t be pulled apart. Based in Quito (UTC−5), working in Spanish or English.

Start a conversation→
GitHub · AndresO7 ↗
  1. 01

    AI systems that survive production

    Agents, MCP servers and gateways, tool governance, context budgets, evaluation that can fail — designed as systems, not demos.

  2. 02

    Cloud architecture on AWS

    Event-driven and serverless, containers on ECS or EKS, the right store for each access pattern — with the trade-offs written down.

  3. 03

    From ambiguity to a phased plan

    Contracts, SLOs, decision records, honest limits and a verification path — and then building it, phase by phase.

  4. 04

    Harnesses for AI-assisted engineering

    Skills, subagent workflows, verifiers and memory, so what agents produce is something your team can review and trust.