AI Solutions Manager

Carlos
Barbosa

I lead AI initiatives end to end — discovery, architecture, build, validation, rollout — with executive sponsorship and a team I'm growing. Currently AI Solutions Manager at Harvard Maintenance.

Team

Leading 2 engineers

Education

M.S. Data Science, UT Austin

View selected work

About

Building intelligent
systems that
deliver impact.

11

Shipped

5

In Production

2

Engineers Led

AI Solutions Manager at Harvard Maintenance, where I own the AI portfolio — setting the roadmap, running discovery with executive sponsors, and staying hands-on in the architecture. M.S. in Data Science from UT Austin.

I take initiatives from the first discovery session through production: framing the problem with directors and C-suite leadership, designing the system, building it, then proving it works before it earns any autonomy. I lead a team of two engineers I'm developing toward full-time roles.

AI Agents & LLM SystemsMulti-agent pipelines, tool use, structured output, prompt versioning
Evaluation & AI SafetyGolden sets, CI eval gates, confidence calibration, shadow-mode validation
ML & Deep LearningPyTorch, Transformers, Scikit-learn, Feature Engineering
Backend & DataPython, FastAPI, SQLAlchemy, SQL Server, Event-Driven Architecture
Cloud & MLOpsAWS Bedrock, Azure, Docker, CI/CD, Model Registry & Versioning

Leadership & Scope

OwnershipAI portfolio — strategy, roadmap, architecture, delivery
StakeholdersC-suite sponsorship (CEO, CFO, COO, CLO) and director-level partners
Team2 engineers, mentored toward full-time roles
DeliveryDiscovery and brainstorming through build, validation, and rollout

Selected Work

Case Studies

01

Enterprise Agent Platform · Program Lead

PythonFastAPIAWS BedrockAngularSQL ServerLLM Evals

Enterprise AI Decision Layer

An AI decision layer over a live service-operations platform, built as an event-driven agent pipeline: it polls system state on a short cycle, assembles the same context a human reviewer would see, classifies risk with an LLM, and fuses that into a composite confidence score. Above threshold it acts through the platform's own permitted-action API — never writing to the database directly, never inferring what it may do. Below threshold it hands a human a pre-filled decision card instead of a blank ticket.

Led it from discovery to approved production program — charter, roadmap, executive stakeholder sessions, and a build-vs-buy analysis that redirected a six-figure external design proposal in-house. Validated observe-only before earning any autonomy: 10 models across 2 providers benchmarked against a hand-labeled golden set, a ≥90% accuracy gate in CI, thresholds derived from Wilson lower-bound agreement curves rather than chosen by hand, and every decision — executed, shadowed, or escalated — written to an auditable ledger with its rationale, evidence, and the policy in force at that moment.

Proprietary — no public link
02

AI Grant Discovery Platform

Next.jsReactTypeScriptPythonAzurepgvectorOpenAIStripeMCP

GrantLens

AI-powered grant discovery platform for US nonprofits. Aggregates and deduplicates federal, state, county, and foundation opportunities (Grants.gov, SAM.gov, USASpending, state portals, and foundations) into one search, ranks them against each org's mission using OpenAI embeddings over pgvector, and predicts eligibility against 174 requirements across 22 funders before you apply.

Live SaaS replacing weeks of manual portal-hopping — with AI matching, a readiness engine, LOI drafting, funder intelligence, and an MCP server exposing 10 tools to Claude.

03

Personal Project · Multi-Agent AI System · Decision Layer Open-Sourced

PythonLangGraphOpenAIAlpaca APIPostgreSQLAzureReact

Argus — Autonomous Trading Agent

An autonomous trading system for US equities, crypto, and options, built as a five-stage agent pipeline — screener, research, signal, risk, execution — with 15+ guardrails that run in two postures: binary rejection, or a sliding scale that shrinks position size instead of refusing outright. A learning subsystem retunes signal weights, thresholds, and regime-specific sizing from realized outcomes, with regime detection, MFE/MAE outcome tracking, P&L attribution, and counterfactual logging underneath it.

I open-sourced the decision layer — the agent pipeline, the guardrails, and the adaptive-learning subsystem — with the brokerage and notification adapters replaced by signature-compatible stubs, so the repo can be read and audited but cannot place an order. The part worth studying is the validation lane: a proposed parameter change becomes a candidate row, runs in shadow against live bars, and only reaches production after promotion and binomial significance gates clear it — rejections recorded, not silently dropped. Published results are measured rather than marketed, including where the system is weak.

04

Financial Analysis AI

PythonOpenAI GPT-4LangGraphChromaDBRAG

PNL Report Agent

AI-powered financial analysis system that automatically identifies anomalies in P&L reports using multi-agent architecture with statistical analysis, vector similarity search, and GPT-4 report generation.

Automated anomaly detection across financial statements with natural language explanations.

05

Full-Stack Nonprofit Platform

ReactAzure FunctionsPythonSQL ServerDocker

TACOLCY CRM System

Full-stack CRM for Belafonte TACOLCY Center nonprofit. Features client intake with tablet kiosk mode, CANS assessments, FNSP eligibility, donor management, and real-time analytics dashboard.

HIPAA-compliant system serving a Miami-based nonprofit with audit logging and role-based access.

06

Regulated Voice AI · Reference Architectures

PythonLiveKitSIPLLMHIPAAPCI DSSTerraform

Enterprise Voice Agents

Two open-source, production-shaped voice-agent reference architectures for regulated domains, on a shared half-cascade LiveKit + SIP pipeline (native audio → text LLM → streaming TTS) with safety guardrails enforced in code, not prompts. PatientLine is a HIPAA-aware patient-access agent (no medical advice, no refill approvals, no PHI before identity verification); WillCall is a PCI-aware ticketing agent whose architecture keeps card data entirely out of the AI environment.

Safety is mechanically provable: release-blocking CI evals enforce zero clinical advice / zero refill approvals (PatientLine) and a Luhn-checked scan proving zero card numbers in any transcript or log (WillCall) — shipped with Terraform IaC, BAA/PCI data-flow mapping, and ADRs defending every decision.

07

Built and Rolled Out

AngularPythonAWSSQL ServerMicrosoft Entra IDSSO

Time & Attendance Platform

An employee-facing portal for time, attendance, and paid-time-off management, replacing manual PTO tracking for a distributed workforce. Angular front end on a Python API, hosted on AWS over SQL Server, with single sign-on through Microsoft Entra ID so it authenticates against the directory employees already use. I developed and implemented it end to end — requirements through deployment — and ran it as a controlled pilot before any wider release.

Piloted with corporate staff, now expanding to branch locations: a staged rollout that used the pilot to settle real workflow and edge cases before scaling, rather than a single launch across every site at once.

Proprietary — no public link

Other Projects

Grant management system for tracking metrics and objectives

AzureStreamlitSQL Server

Imitation-based AI agents in simulated environments

Deep LearningImitation Learning

ELECTRA transformer robustness in NLI tasks

ELECTRANLPEnsemble Learning

Customer purchase behavior prediction for marketing optimization

StreamlitMLPython

Credentials

Education &
Certifications

M.S. Data Science

University of Texas at Austin

2024

B.S. Industrial Engineering

FAESA

2016

Data Science & AIPython, ML, Deep Learning
10+
Business IntelligencePower BI, Excel, SQL
8+
Cloud & DevOpsAzure, AWS, Git
5+
Agile & ManagementScrum, Sprint Planning
3+
View all on Credly

Contact

Let's build
something
together.

Open to opportunities in AI implementation and data science — including contract and advisory work through my firm, Kojo Analytics. If you have a problem worth solving, I'd like to hear about it.