Dennis Salguero

AI Engineer

+1-725-777-6815

[email protected]

  • Experience working in regulated environments

  • Have led teams of up to 15 data scientists

  • Open to relocation

  • US Citizen

Multi-Modal AI Coach


Problem

Career coaching is a high-consideration purchase. Prospects cannot evaluate whether a coach's tone, judgment, and advice fit them without experiencing it first, and a single discovery call is a high-friction commitment for both sides. The result is a long evaluation cycle and lost leads who never reach the call.


What I Built

A production coaching agent deployed across four channels (web, DMs, SMS, and inbound voice) sharing one RAG and prompt infrastructure. Built on FastAPI with Postgres and pgvector backing a custom corpus of book and podcast transcripts, embedded with Voyage AI voyage-4-large at an empirically calibrated cosine-distance cutoff. The runtime routes across three Anthropic Claude models per job: Opus 4.7 for primary coaching, Sonnet 4.6 for mobile channels, and Haiku 4.5 for post-session structured lead scoring, selecting per call for latency and cost. The coaching system prompt is versioned across 13 revisions and four channel variants, with conversation-act directives composed per turn. Voice uses an ElevenLabs clone over Twilio TwiML. Approximately 24,000 lines of Python across 149 files with 43 test files.


Outcome

More than 1,000 conversations have occurred across all channels resulting in more CTA presentations and new business. Clients have provided positive feedback on the human-like nature of the bot and that is does not give generic advice.

High Volume Social Media Publishing Engine


Problem

Backlog of more than 5,000 social media posts that need to be organized and published on an ongoing basis


What I Built

A Python/Flask application backed by SQLite, with direct API integrations against Meta Graph v22 (Instagram, Facebook, Threads), YouTube Data API v3, LinkedIn REST v2 (personal and Company, including chunked video upload), X/Twitter, and the Threads API. Captions are generated by an OpenAI pipeline that runs Whisper for transcription and gpt-4o-mini for caption generation, with all 4,900+ videos preprocessed into a SQLite cache so per-post inference cost at publish time is zero. Publishing is driven by a weighted-random selection engine with per-platform cooldowns, least-recently-posted tiebreaking, dayparting, and automated graduation between content buckets. Approximately 14,000 lines of Python.


Outcome

More than 5,000 posts have been published on an automated basis, running 24/7, resulting in account growth and new business

Multi-Subsystem Hospitality Pipeline


Problem

Tourism & hospitality face a lot of competition and the ability to answer customer inquiries is key. Human agent response is not always available quickly


What I Built

A four-subsystem Python 3.12 architecture (crawler, video renderer, Instagram publisher, DM concierge) with a strict file-based boundary and per-subsystem CLI verb namespacing, each independently testable and deployable. The DM concierge runs on Anthropic Claude Sonnet 4.6 with native tool use inside a bounded three-turn lookup loop, plus a sentinel-tag protocol for out-of-band side effects (email capture, operator handoff). During build I resolved a one-month Meta integration blocker by establishing that Instagram DM access requires the separate Instagram Business Login product with its own credentials and graph.instagram.com host, distinct from Facebook Login. Approximately 9,700 lines of Python with 334 tests.


Outcome

Currently in production, serving clients in Las Vegas