Juan David Gomez Villalba

Senior Software Engineer (Mar. 2025 – Present) · Software Engineer (Nov. 2024 – Mar. 2025)Toronto, ON (Remote)

• Architected and operate the AI lead-signals engine. Agents search 22K companies a day and turn 1.17M search results into 8,660 sales signals in 13.5 hours, at $31 per 1,000 signals.
• Cut AI spend 90% ($100K+/yr, est.) while scaling 4x. 70K → 280K articles/day and 1 → 4 regions. Deterministic prefilters send under 2% of search results to an LLM; spend circuit breakers and per-signal cost tracking cover the rest.
• Made the AI trustworthy. Built the LLM evaluation and human-review platform that gates 5,900 CRM updates a day, with per-client kill switches; redesigned the evaluator to cut false “already in CRM” claims 89% → 3%. Added AI decision gates that refuse instead of guess.
• Led the fixes on live client data. Across 156K relationship records, cut wrong-owner records 6.5% → 1.0%; removed 61K duplicate contacts with zero measured identity-safety violations.
• Shipped the CRM and lead-gen features users open daily. Smart Outreach (finds the right decision-maker for each signal, drafts the email, one-click send), an Outlook-simple outreach screen non-technical users adopted, delegated sending, job-change alerts, AI reply classification, and the first CRM Cleanse on HubSpot.
• Took the AI email platform from prototype to production. FastAPI, Celery and Redis; CrewAI, then Autogen with Perplexity research. Launched Deep Research and Keep In Touch agents, extending the product from sales into client retention.
• Built it to keep running. Jobs resume after crashes, nothing is double-counted, and every failure is visible end to end across three repos (BullMQ, Celery, KEDA on Azure).

Forward Deployed Engineer (Contract), working directly with the owner · 32-room hotel, CuraçaoRemote

• Gave a small business its own AI operations stack. The owner asks about bookings in plain English (AI connected to Cloudbeds, their booking system) and gets a 7am daily brief: arrivals, balances to chase, 14-day occupancy.
• Automated pricing and bookings. They run unattended; a replay over 131K pricing decisions reproduces the incumbent revenue manager exactly; 20 months of sales data reshaped the menu. Shipped the website.

Software Developer, Blockchain Security (Part-time)Remote

• Real-time fraud and exploit detection. 800K–1.2M datapoints/day with zero data loss; 80% faster processing; 8+ ML alerting models in Docker.

BSc Computer Science (Machine Learning & AI) · First Class HonoursLondon, UK