Skip to content
    Axel Cuevas
    • Home
    • Blog
    • Projects
    • About
    • Resume
    🇺🇸
    Let's talk

    AxelCuevas

    Backend & AI engineer. Warehouse systems, payment rails and agents with permissions, built to stay up.

    axel@devhub~
    roleAI Engineer @ Leadsales
    focusagents · payments · warehouses
    basedSanto Domingo · 17:51 AST
    commits · 7d32
    stackTypeScript · Python · NestJS
    Scroll

    Move the light over the setsTap to aim the light

    Most of what I build runs where nobody is watching: in a warehouse measuring pallets, between a checkout and three payment gateways, in a Slack channel where an agent waits for permission. Today I do it as an AI engineer at Leadsales.

    Selected work

    Things I shipped.

    All eleven ↓
    01Rivero · 2026
    In production

    Pallet Dimensioner

    A depth camera, a Python point cloud, a pallet measured in under a second.

    $25k+
    saved per station
    <1 s
    per pallet
    02Rivero · 2026
    Internal

    Permission-Gated AI Orchestrator

    An agent that runs backend operations from Slack, and can only touch what it was allowed to.

    40%
    of low-complexity tickets closed on its own
    100%
    of capabilities behind a scope
    03Rivero · 2025 — 26
    In production

    Warehouse Platform

    Receiving, cycle counts, reporting and the semantic layer under one monorepo.

    +80%
    inventory traceability and accuracy
    ~20%
    more revenue in under two months
    04IQ BIAI Solutions · 2025
    In production

    One interface over Cybersource, CardNET and Azul.

    3 → 1
    gateways behind one interface
    -70%
    integration complexity
    More work08
    Stack

    What I actually reach for.

    Hover a star · tools that ship together are wired
    // Stack chart
    Epoch J2026.7 · 50 stars · 6 constellations
    Hover a star · click to pin
    bigger = used more
    Activity

    How I build.

    GitHub · contributions · last 12 months2,860 total
    MonWedFri
    OctNovDecJanFebMarAprMayJunJulAugSepOct
    Agent sessions per week · Claude Code + Cursor · since 2026-07-16
    0 contributions in the last yearLess More
    When I prompt · weekday × hour (Santo Domingo)1,839 prompts
    SunMonTueWedThuFriSat
    061218
    Agents · Claude Code + Cursoras of 2026-10-02
    Written by agents?
    99.9 % of lines
    24,643 lines across 104 tracked commits in Cursor · every diff reviewed by me
    Delegation?
    14 subagents
    per session · 952 runs across 67 Claude Code sessions
    Hands-off?
    92 % unattended
    of prompts ran in auto or bypass mode · 101 stopped to ask me first
    Which models?
    Opus 5CCOpus 5.5CCGrok 4.6CU
    picked per job · Opus 5 + Opus 5.5 in Claude Code · Grok 4.6 in Cursor
    How much?
    376 h
    67 active days · 75 sessions · 8 projects · since 2026-07-16
    When?
    Peak 11 AM
    Thursdays · 14 % after 10 PM · best streak 23 days
    Session ledger · Claude Code · latestall weeks
    2026-10-02second-brain · Opus 5.550 min8 p
    2026-10-01misc · Opus 5.587 min3 p
    2026-09-30portfolio · Opus 5.5395 min65 p
    2026-09-30misc · Opus 5.5808 min259 p
    2026-09-30nest-base · Opus 5.55 min1 p
    2026-09-28rivero-track · Opus 5.51 min2 p
    2026-09-27portfolio · Opus 5.518 min6 p
    2026-09-27misc · Opus 5.51284 min86 p
    Experience

    Where the work happened.

    CV (PDF)
    • Cut the production Lead Agent's inference cost 19x in the first week by re-evaluating model selection and migrating the agent to a far cheaper model.
    • Rewrote the agent's conversational behavior to strip generic AI phrasing, so replies read like a human sales rep and sell more naturally in customer conversations.
    • Own the GenAI architecture behind the agents that acquire, qualify and serve Leadsales' customers: RAG over OpenSearch and Pinecone, LLM-based lead classification, multi-format knowledge bases (PDF, Excel, OCR) and human–agent handoff, with cost, latency and quality tracked in production.
    • Build the team's AI-native workflow: packaged skills, deterministic guardrails around agentic work (hooks, CI gates), documented architecture decisions and mentoring on LLM patterns.
    Remote · stack
    PythonTypeScriptclaudePineconeopensearchLangGraphMCP
    • Engineered a proprietary 3D pallet dimensioning system (hardware and software) using Orbbec ToF depth cameras, WebSockets and a custom Python 3D extraction algorithm, designed to bypass a competitor's US patent, saving $25k+ per station and drawing B2B licensing interest from the DR's largest supermarket chain.
    • Architected Orchy, a permission-gated AI orchestrator built on OpenClaw and Hermes primitives with strict security scopes across all MCP integrations, so the AI safely executes backend operations (such as Google Workspace provisioning) from Slack. Its self-learning loop resolves 40% of low-complexity engineering tickets while preventing unpermissioned LLM hallucinations.
    • Built Rivero's MCP server inside the NestJS API, with an OAuth 2.1 authorization server per the MCP 2026-07-28 spec, so anyone can connect Claude, ChatGPT or any MCP client to the WMS under their own permissions to query warehouse data, build and schedule reports and onboard users, with tools gated per user and every call audited.
    • Owned a mobile-first Warehouse Audit System end to end, replacing blind counts with a risk-based smart routing algorithm and BullMQ for asynchronous pallet re-scoring and audit logging; resolved database-level anomalies (multi-location ghost pallets), raising inventory traceability and accuracy by 80%.
    • Architected a semantic reporting engine on Cube.js over a PostgreSQL read replica with pre-aggregations, and an interactive AG Grid + ECharts front end with SKU-level drill-downs and actionable views for bulk liquidations and offers, increasing revenue by ~20% in under two months.
    Remote · Santo Domingo, DR / Miami, FL · stack
    TypeScriptPythonNestJSNext.jsReactPostgreSQLRedisBullMQClickHouseMCPDockerLinux
    • Architected a multi-tenant payment API in NestJS that unified 3 major Dominican gateways (Cybersource, CardNET, Azul) under a single interface, reducing client integration complexity by 70%.
    • Engineered a robust security architecture: an HMAC-based API Key system and an end-to-end tokenization flow with Cybersource Flex Microform, reducing PCI DSS compliance scope to SAQ A.
    • Designed a highly available webhook dispatch system on BullMQ with Redis, ensuring 100% delivery reliability with exponential backoff for critical payment and refund notifications.
    Remote · stack
    NestJSTypeScriptNode.jsPrismaRedisBullMQPostgreSQLDocker
    • Founded and architected DeUna, an autonomous conversational AI sales closer for LATAM SMEs, operating natively inside WhatsApp to automate DMs, scheduling, and high-ticket sales.
    • Engineered a low-latency voice pipeline using Whisper v3 Turbo and ElevenLabs to clone business owners' voices, processing heavy regional dialects for auditory authenticity.
    • Implemented a bounded-haggling algorithm inside the RAG engine, letting the AI negotiate dynamically with sales psychology while mathematically protecting the hidden price floor.
    Remote · stack
    PythonTypeScriptFastAPILangGraphPineconePostgreSQLRedisDockerAWS
    • Developed a real-time bidirectional audio layer using WebSockets and the Gemini Live API, reducing voice response latency by 60%.
    • Engineered robust data pipelines with Dagster, dbt and ClickHouse, applying A/B testing and transformation models.
    Remote · stack
    PythonFastAPIGeminiGoogle ADKMCPDagsterdbtClickHouseDocker
    Contact

    Have a system that needs building?

    axeljcuevast@gmail.com
    Santo Domingo · · open to select freelance
    Keep exploring
    AboutWho I am and how I workProjectsCase studies, systems and codeResumeExperience, stack and CV
    © 2026 Axel CuevasGitHub · LinkedIn · X · Instagram · CVBuilt with Next.js · three.js · GSAP
    Writing

    Latest post.

    All posts
    October 6, 2026 · MCP · Model Context Protocol

    The Tool List Is the Permission Check: Building a Per-User MCP Server with OAuth 2.1

    How I put a warehouse system inside Claude and ChatGPT without giving any model more access than the person using it: one MCP server inside the existing API, tools gated at registration, an audience check on every token, and a write ledger for agent retries.

    Read