Case studies
Systems that touch real money.
Each one runs today, in front of real customers. No mockups, no "coming soon."
Sofía — the AI salesperson that never sleeps
A bilingual (EN/ES) sales agent for a U.S. car dealership: answers, qualifies and books appointments across web chat, WhatsApp and voice. Every outbound message passes guardrails plus an LLM-as-judge verification layer before a customer sees it. Playbook engineered from 1,200+ real sales conversations; voice runs on latency-routed models.
→ live at cylinderautosales.com · answering customers right now
The loop advertisers never close
Lead phone number → CRM pipeline stage → Meta Conversions API purchase event. The ad algorithm learns from closed sales instead of clicks. Built the same discipline inward: a sales radar that judged 4,088 conversations and proved the CRM was seeing ~1% of real sales activity.
→ ad spend now optimizes on revenue, not vanity metrics
Agency OS — prospecting at machine scale
Scored and LLM-enriched 11,800+ dealerships into 2,135 qualified prospects, each with a personalized brief and hook. An autonomous AI analyst reads sales activity every morning and reports "the 3 actions worth money" via Telegram — on a cron, no humans required.
→ zero manual research · runs daily unattended
An Android device farm with bank-grade discipline
Fleet of physical phones (ADB/uiautomator2) publishing dealer inventory, with live telemetry dashboard, Telegram control bot, persistent per-day rate limits and fail-closed safety gates. Hardened through multi-agent adversarial code audits — 14+ real bugs found and fixed.
→ fail-closed by design · full unit-test suite
Trading systems that survive their own hype
Recovered a client's NinjaTrader 8 algorithmic system from a dead VPS, extended it with an ONNX ML inference gate, and built a research pipeline with pre-registered out-of-sample validation — the discipline that kills false edges before they burn capital.
→ client system recovered, extended, and validated
The Office — an operating system for AI agents
A personal command center where a fleet of AI agents — sales, prospecting, QA, orchestration — is visualized as a live pixel-art office: every character on screen is a real bounded workflow with state, run history and an accountable owner. A General Manager agent coordinates the roster and temporary sub-agents spawn on demand, do their slice of work, and disappear. The part that surprises engineers: it runs 100% local on a Claude subscription — no paid API keys, works fully offline, external cost $0.00.
→ runs on a subscription, not an API budget · offline-ready · one person, a whole team on screen
How I work
A running system in 30 days.
Not a slide deck. Not a roadmap. A system, in production, with a measurement loop.
Map & strike
I find your highest-leverage manual workflow and ship a working automation against it. You see running code in the first week.
The agent
An AI agent — chat, voice or back-office — integrated with your real stack (CRM, ads, webhooks), with guardrails and evaluation, running in staging.
Production + proof
Deploy, plus a measurement loop so what it earns or saves shows up in numbers you can read. Then we decide what's next.
Work with me
Your AI engineer, on demand.
Available now for contract. Spanish native / English fluent — I build and sell in both. US Eastern time.
Availability
- 20–40 hrs/week, start this week
- Remote (US Eastern), bilingual EN/ES
- Solo or embedded in your team
What I ship
- AI agents (chat / voice / back-office)
- LLM pipelines, CRM & ads integration
- Guardrails + LLM-verified outputs
My guarantee
- First working deliverable inside 30 days
- Weekly shipped increments, always measured
- Or we don't continue
Stack
Tools follow the problem.
What's currently in production.
Open source: video-grabber (MV3 + ffmpeg.wasm) · seevideo (video → LLM-readable)