boscojacinto

With several years of hands-on applied AI experience and a proven track record of architecting, building, and deploying
production-grade AI applications, I am excited to bring my expertise in AI agents, multi-agent workflows, federated learning, and generative systems to your team.

In my current role as AI Lead at AuxoAI (Nov 2025 – present), I have delivered two high-impact
enterprise solutions: AEC Quantity Take-Off Automation: Developed a multi-agent system that automates quantity
take-offs from highly variable civil construction drawings (storm-water drains, water lines, sewer lines). Faced with noisy PDFs containing overlapping utilities, annotations, symbols, and multi-contractor layouts, I took an AI-agent-first approach backed by rapid advances in vision
reasoning models. I designed a novel Filter → Generate → Ground pipeline: LLM-driven filtering based on PDF metadata (line thickness, tags, etc.), pipe isolation using a specialized image generation model (nano-banana) + potrace vectorization, and vector clustering with dynamic tolerances for grounding. This balanced generalization and accuracy effectively. I orchestrated ~12 specialized agents with careful context distillation and built a custom
evaluation pipeline using Langfuse, golden datasets, vector grounding, and CV template matching to measure recall, precision, and symbol accuracy. To increase stakeholder confidence, I introduced a Human-in-the-Loop (HITL) feature allowing users to validate and correct counts and pipe lengths — turning temporary model limitations into
a collaborative strength.

Clinical Trial Budget Automation: Created an automated workflow for a leading cancer research institution using semantic mapping and tool-augmented agents. Instead of relying solely on cosine similarity, I leveraged the model’s native trained knowledge of clinical trial terminology combined with tool calls, delivering a clean single-agent solution that was later extended to other data migration tasks.
I approach applied AI problems by first identifying workflow invariants, setting clear generalization/accuracy targets, breaking complex tasks into sub-problems, and drawing
deliberate boundaries between heuristic logic and LLM-driven loops. I also actively identify missing pieces — such as tacit knowledge gaps — which led me to champion declarative AI workflows and continuous data gathering for fine-tuning specialized SLMs. This two-model strategy (foundation model + domain SLM) was validated through experiments and presented to leadership, resulting in meaningful shifts in company direction.Complementing this enterprise experience are two flagship projects I built from scratch:
Restaurant-FL (Mar–Jul 2025): A privacy-preserving federated learning platform using FlowerAI. It extracts customer features via natural language, builds heterogeneous graphs with Graph Transformers, and trains local + global models. Integrated Gemma-3:4B (Ollama), Graphiti, and Tendermint consensus.

Headconn (Aug 2025): A multi-agent image generation system (Imagine → Reflect → Composite) powered by Grok-4 structured outputs for creating pop-culture mash-up images
using advanced 2D manipulation tools. These experiences highlight my strengths in rapid research-to-prototype cycles, rigorous custom evals, agentic design, and correcting implementation edge cases through deep debugging (e.g., understanding library-specific data formats like PyMuPDF vector schemas). I thrive in collaborative settings — setting technical direction, delegating based on team strengths, and jumping in to unblock others while sharing strategic insights with leadership.

I seek challenging problems, compelling product roadmaps, and environments filled with talented people who celebrate experimentation in an open, friendly culture. Your work aligns closely with what I’m looking for in my next role.