I build and deploy AI-powered software directly with customers — from discovery to production.
_{US Citizen · Washington DC-Baltimore Area · Remote · 2-Week Notice · Open to Full-Time & Contract}

I'm a Forward Deployed Engineer: I embed with customers, map their workflows, and ship production software that solves them. 9+ years shipping TypeScript/React/Next.js, Python/FastAPI/Node.js, and PostgreSQL — from fintech (SoFi) to government/defense (General Dynamics IT). Now at CalendarFuel, I build and run the LinkedIn outreach platform end-to-end — warm signal detection, LLM ICP scoring, personalized sequences, and autopilot booking — for real customer environments.

Currently in production: ~50 customers and campaigns across 10 companies · workflows processing up to 500K prospects/records per month · ~1 week faster time-to-close through automated research, personalized outreach, and sales intelligence.

I ship with Cursor, Claude Code & Codex, and GitHub Copilot daily — using GPT-4o, GPT-4o-mini, Claude 3.5 Sonnet, and Claude 3 Opus to accelerate architecture, not replace judgment. Every AI system I ship comes with eval gates and a rollback path.

Jan 2024 – Present · Remote

• Build and deploy production AI-powered software for a growing B2B SaaS company, working directly across engineering, product, sales, customers, and internal stakeholders.
• Translate customer workflows and business requirements into shipped features, prototypes, integrations, and production-ready solutions.
• Design integration architecture and data flows across CRM, enrichment, sequencing, communication, and internal systems — REST APIs, webhooks, async workflows, structured reporting.
• Build LLM-powered workflows, agents, dynamic personalization, and intelligent automation with TypeScript, React, Next.js, Python, and FastAPI.
• Support ~50 customers and campaigns across 10 companies, including client teams of 5–10 running active campaigns.
• Operate 20–30 active LinkedIn profiles for internal sales teams; support workflows processing up to 500,000 prospects/records per month.
• Cut time-to-close by ~1 week via faster data flow, personalized outreach, automated follow-up, and clearer sales intelligence.
• Own production troubleshooting, AI output evaluation, and customer feedback loops across the full SDLC.

Jan 2022 – Dec 2023 · West Palm Beach, FL

• Developed secure full-stack web applications for federal government clients with Next.js, React, TypeScript, Node.js, and Python (FastAPI/Flask), meeting strict security compliance standards.
• Implemented RESTful and GraphQL APIs integrated with AWS services and relational databases — 99.9% uptime for mission-critical systems.
• Led code reviews and mentored junior developers; introduced automated testing with Jest and Cypress and CI/CD pipelines with GitHub Actions.
• Migrated monolithic systems to microservices architecture using Docker and Kubernetes.

Jun 2019 – Dec 2021 · Remote

• Built Python (FastAPI) backend services and async data pipelines powering marketing and sales operations — ingesting, transforming, and syncing customer data between HubSpot, Salesforce, Dotdigital, and internal warehouses.
• Developed REST API connectors and webhook handlers keeping marketing automation and CRM in sync, so lead scoring, segmentation, and attribution data stayed accurate.
• Implemented validation, deduplication, and enrichment logic for clean records flowing into sales and marketing tools.
• Partnered with marketing ops and sales to debug sync issues, optimize API throughput, and build internal tooling for pipeline data health.
• Designed backend endpoints and aggregation jobs feeding campaign, conversion, and funnel metrics into leadership dashboards.

Jan 2017 – Jun 2019 · Remote

• Designed automated test suites for web applications using TypeScript, Cypress, and Playwright — E2E, integration, and regression coverage across front-end and back-end.
• Integrated automated testing into CI/CD pipelines, catching defects earlier and improving release reliability.

Education: Washington University of Science and Technology — Full-Stack Web Development Bootcamp (Cybersecurity, Software Engineering, Project Management)

Every engagement runs the same loop: audit → eval → deploy → observe. These docs are the artifacts, organized by stage.

Audit — map how the work actually happens before touching it.

• Decomposition Case Studies — breaking down ambiguous customer problems into shipped systems
• FDE Audit: Operating Map — a week-one audit deliverable: documented vs. real process, exception inventory, ROI priority matrix

Eval — decide where intelligence belongs, then gate it.

• Production AI Evaluation — AI evals, latency budgets, and safety in production
• Where Intelligence Belongs — the deterministic vs. LLM vs. human-in-the-loop decision framework

Deploy — land on the customer's existing stack, with graduated autonomy.

• Customer Reference Architecture — AI-native GTM architecture for B2B SaaS
• Deployment & Adoption — no forced migration, shadow-mode rollout, de-risking the internal champion
• GCP + Vertex AI RAG Agent — retrieval-augmented agent with evals, deployed on GCP

Observe — audit trails, failure-mode reports, then the loop runs again.

• Production Incident Post-Mortem — rate-limit failure in an AI pipeline, root cause to resolution
• SQL & Data Modeling Deep Dive — query patterns and schema design from production systems

All three systems below are live in CalendarFuel (calendarfuel.co) — the LinkedIn outreach platform I build and operate for ~50 customers across 10 companies. The decomposition case studies in my FDE Portfolio are two client deployments of this platform written end-to-end — the metrics here and there are the same systems at different altitudes.

Customer problem: Clients came to us after burning out on cold outreach — static lists, 2% reply rates, and restricted LinkedIn accounts. Deployed solution: a warm signal engine that only contacts people already showing intent — no cold lists.

The platform pulls profile viewers and post engagers from clients' own LinkedIn accounts via Unipile every 2 hours, and layers in hiring-intent and ICP-matched profile search via Apify every 4 hours. Every signal lands in a Supabase lead inbox before any outreach happens.

Clients reply to outreach because it's never really cold — one prospect's literal reply: "how did you know I was looking at this?" The system found the signal and wrote the message.

Customer problem: Every client has a different ICP, and manual qualification doesn't scale across ~50 concurrent campaigns. Deployed solution: an LLM scoring pipeline that runs every 30 minutes against every new lead.

Each lead is scored 0–100 against that specific client's ICP — title, seniority, company size, funding, tech stack, recent activity. Score ≥ 60 routes to campaigns; below 60 is disqualified before it ever wastes a send. Open profiles (LinkedIn gold badge) get a +3 bonus and route to the InMail track instead of connection requests. Primary model via OpenCode with an automatic OpenRouter fallback, so a provider outage never stalls the pipeline.

Multi-tenant by design: scoring runs per-client ICP definitions from one shared pipeline, with per-campaign sender isolation so one client's backlog never starves another's queue.

Customer problem: Every automation tool clients had tried before got their accounts restricted within a month. Deployed solution: a webhook-driven campaign state machine that behaves like a careful human, not a bot.

Connection requests go out at human-like pace with daily limits and warm-up scheduling. DMs are hard-gated behind the connection.accepted webhook — a value DM is only scheduled after LinkedIn confirms the connection, then sent 2 hours later. Follow-ups escalate to an AI voice note (ElevenLabs TTS delivered through Unipile) instead of another text wall. Health monitoring watches for rate-limit risk and shadow-ban signals before LinkedIn flags anything.

Interested prospects book directly on the client's calendar via Calendly, with a Slack notification to the owner. Clients show up to meetings; the platform handles everything between intent and booking.

flowchart TB
    subgraph Client["Client Experience"]
        UI[Next.js 14 Dashboard]
        INBOX[Unified Reply Inbox]
    end

    subgraph Core["Core Platform — Netlify"]
        API[Next.js API Routes]
        DB[(Supabase Postgres<br/>leads · messages · campaigns · clients)]
        CRON[Scheduled Functions<br/>message sender 5min · engagement sync 2h<br/>ICP scoring 30min · lead sourcing 4h]
    end

    subgraph External["External Services"]
        UNI[Unipile<br/>LinkedIn automation]
        APF[Apify<br/>hiring intent · profile search]
        LLM[LLM scoring + message generation<br/>OpenCode → OpenRouter fallback]
        EL[ElevenLabs<br/>AI voice notes]
    end

    subgraph Platform["Auth · Billing · Alerts"]
        CLK[Clerk]
        STP[Stripe]
        SLK[Slack]
    end

    UI --> API
    INBOX --> API
    API --> DB
    API --> CLK
    API --> STP
    CRON --> DB
    CRON --> UNI
    CRON --> APF
    CRON --> LLM
    CRON --> EL
    CRON --> SLK
    UNI -->|webhooks: connection.accepted, replies| API

flowchart TD
    A[Warm signal detected<br/>profile view · post engager · hiring intent] --> B[LLM ICP scoring 0–100<br/>every 30 min]
    B --> C{Score}
    C -->|≥ 60| D[Qualified → campaign]
    C -->|< 60| E[Disqualified]
    C -->|Open profile +3| F[InMail track]
    D --> G[Connection request<br/>human-paced, daily limits]
    F --> G
    G --> H{connection.accepted<br/>Unipile webhook}
    H -->|accepted| I[Value DM<br/>scheduled +2h]
    I --> J[Follow-up sequence]
    J --> K[AI voice note<br/>ElevenLabs → Unipile]
    K --> L{Reply}
    L -->|interested| M[Autopilot booking<br/>Calendly + Slack alert]
    L -->|objection| N[AI objection handling<br/>+ re-engagement]
    N --> M

Daily stack: Cursor (IDE) · Claude Code & Codex (refactors) · GitHub Copilot (boilerplate)
Models: GPT-4o · GPT-4o-mini · Claude 3.5 Sonnet · Claude 3 Opus · OpenRouter

Workflow: Customer discovery → natural language spec → AI-accelerated implementation → human-owned architecture, error handling, and production reliability. Every engagement runs the same loop — audit → eval → deploy → observe — and the loop is the deliverable as much as the code.

I deploy on GCP, AWS, and Vercel — infrastructure decisions are driven by customer constraints, not preference.

• Forward Deployed Engineer — embedded with customers, shipping AI-native systems from discovery to production

Must-haves: Remote-first, US-based, small high-velocity team
Start: 2-week notice

📧 madelynreyes2026@gmail.com
🌐 Washington DC-Baltimore Area (Remote) · US Citizen · No sponsorship needed