Stephen Bonifacio

Manila, Philippines

  • AI Engineer building production LLM systems and AI agents at enterprise scale
  • Shipped multi-tenant AI platforms serving 4,000+ users across a large conglomerate
  • Built an agent harness that runs org skills in isolated cloud sandboxes, with any LLM provider
  • End-to-end ownership: agent loops, RAG, tooling, frontend, infra, and observability
  • 10+ years designing and delivering mission-critical enterprise software
Stephen Bonifacio

experience

AI Architect

JG Summit Holdings Inc — Manila, Philippines
January 2023 – Present

Led the design and delivery of the group's production AI systems, from a multi-tenant LLM platform serving 4,000+ users across business units to the infrastructure and tooling that run it.

  • Cut HR and IT support response times from days to seconds with an enterprise AI agent integrated with ServiceNow and SAP HCM
  • Architected the modular platform behind it (bring-your-own LLMs, vector DBs, tools, and prompts) and extended it into Microsoft Teams
  • Built a sandboxed agent harness that runs org skills on any LLM provider, and a remote MCP server giving the agent per-user access to SAP HR data
  • Replaced static customer experience surveys with an AI interviewer that captures the reasoning behind each rating
  • Gave non-technical admins and executives control and visibility: an admin console for the knowledge base and conversations, plus automated monthly usage reports
  • Owned observability (Langfuse) and CI/CD for every AI application across dev, staging, and production

Python • TypeScript • FastAPI • Next.js • Azure AI Foundry • Azure AI Search • Azure Container Apps • Cosmos DB • PostgreSQL • MCP • Langfuse

projects

Enterprise AI Agent

Designed and shipped a production-grade, multi-tenant LLM platform that augmented HR support workflows across multiple companies

  • Modular, bring-your-own architecture (LLMs, vector DBs, tools, prompts)
  • Async FastAPI backend optimized for concurrent RAG + tool execution
  • Bespoke 'agent loop' implementation for full control and extensibility (parallel tool calls, custom retry logic, dynamic prompt injection, etc)
  • Multi-tenant Next.js app with real-time streaming, branding, and isolation
  • Enterprise-grade security: Entra ID SSO, 2FA, JWTs, encrypted PII storage
  • Integrated ServiceNow (RAG + tickets) and SAP HCM (employee data access)
  • Full observability with Langfuse (latency, tokens, evals, prompt versions)

Python • TypeScript • FastAPI • Next.js • Azure AI Search • Azure CosmosDB • Redis • Langfuse

Sandboxed Agent Harness for Org Skills

Built an agent harness that runs curated organizational skills, with the agent loop on the server and all code execution isolated in a per-session Azure Container Apps sandbox (hardware-isolated microVMs)

  • Bespoke agent loop with tool routing, context compaction, error recovery, and cancellation-safe tool calls
  • Every shell and file operation runs in the user's own sandbox, never on the harness host; workspaces persist on a per-user volume
  • Sandbox lifecycle management: pre-warmed spares, idle stop with transparent resume, and a reaper for expired sessions
  • Progressive skill loading: the catalog sits in the system prompt, full instructions and scripts load on demand, with org-based access control and admin publishing
  • Multi-provider LLM layer (Claude on Azure AI Foundry, OpenAI, and others) driven by a config-only model catalog with cross-vendor fallback
  • Append-only conversation log with a projected message view, enabling non-destructive compaction, rewind, and feedback
  • MCP connectors, Entra ID auth, and Langfuse tracing and prompt management, plus a Next.js chat UI with real-time SSE streaming

Python • TypeScript • FastAPI • Next.js • Azure Container Apps Sandboxes • Azure AI Foundry • PostgreSQL • Azure Blob Storage • MCP • Langfuse

Observability Platform for Admins

Built an internal observability and control plane enabling non-technical admins to operate, inspect, and improve production LLM systems

  • Created document management interface enabling non-technical users to manage RAG knowledge base
  • Built document ingestion pipeline for ServiceNow KM articles with automatic chunking and embedding
  • Built conversation explorer with message logs, knowledge gap detection, and RAG pipeline inspection for admin oversight
  • Implemented usage analytics dashboard tracking token consumption, tool usage, and peak engagement patterns
  • Added ticket integration tracking and dark mode support

TypeScript • Next.js • shadcn/ui • Azure AI Search • Azure CosmosDB • Redis

Microsoft Teams Integration

Extended the core enterprise LLM platform into Microsoft Teams, enabling employees to access the same AI capabilities directly within their primary collaboration tool

TypeScript • Express.js • Agent Framework • Azure Bot Service

Demo & details
  • Implemented a Microsoft Teams bot as an additional frontend channel for the existing LLM platform
  • Reused shared backend services, RAG pipelines, tools, and security model
  • Built the bot using Microsoft Bot Framework and Express.js
  • Integrated with Teams messaging APIs for real-time, bidirectional communication
  • Implemented adaptive cards for structured, interactive responses in chat
  • Deployed and managed via Azure Bot Service alongside the core platform

Usage Analytics & Reporting Engine

Designed an automated analytics and reporting engine providing executives and stakeholders with ongoing visibility into AI platform usage and value

Python • Azure Functions • CosmosDB

Demo & details

Click to view

  • Built a scheduled reporting system delivering monthly AI usage reports via email
  • Aggregated and deduplicated conversation data across all tenants and applications
  • Generated insights covering cost trends, usage patterns, knowledge gaps, and top users
  • Implemented fault-tolerant pipelines with automated error handling and retries
  • Enabled month-over-month trend analysis to guide platform investment decisions

AI-Powered Customer Experience (CX) Interviewer

Replaced a static customer experience survey with a conversational AI interviewer that probes each rating with dynamic follow-ups, capturing the qualitative "why" that traditional questionnaires miss

TypeScript • Next.js 16 • React 19 • Azure OpenAI • Langfuse • Azure CosmosDB • Azure Easy Auth

Demo & details
  • Two-phase flow per question (score reasoning → what good looks like), independent state each
  • Bounded AI: opener + 3 follow-ups, then graceful close — via versioned Langfuse prompts
  • SSE streaming over Azure OpenAI SDK with a custom React hook for lifecycle management
  • Full observability with Langfuse via OpenTelemetry (traces, tokens, prompt versions, sessions)
  • Structured JSON logging with PII redaction, correlation IDs, and dev/prod level gating
  • Server-side CosmosDB persistence (per-question docs, session-partitioned) behind Azure Easy Auth

Remote MCP Server for SAP HCM

Built a Model Context Protocol (MCP) server exposing SAP HCM employee data to the enterprise AI Agent for conversational access to personal, employment, timekeeping, and payroll records

Python • FastMCP • httpx • Tenacity • Azure App Service • Azure Easy Auth • Azure API Management • SAP HCM

Demo & details
  • Single discriminated MCP tool covering four SAP HR domains — keeps the LLM's tool surface small and selection accurate
  • Async FastMCP server over httpx with tenacity retries + exponential backoff for flaky SAP responses
  • Per-domain parsers turning SAP's flat key-value payloads into structured, LLM-friendly output
  • Azure Easy Auth flows the caller's identity into the SAP lookup, so each user sees only their own HR data
  • Deployed on Azure App Service (streamable-http)
  • Per-request correlation IDs and header-based client auth for enterprise traceability

DevOps & Infrastructure

Built and operated the CI/CD and cloud infrastructure underpinning all production AI applications

Azure DevOps • Azure Pipelines • Azure App Service • Git • Bash

Details
  • Designed CI/CD pipelines in Azure DevOps for the backend API, web app, Teams bot, and admin app
  • Automated build, test, and deployment workflows across dev, staging, and production
  • Deployed services to Azure App Service with environment isolation and secrets management
  • Standardized release processes to support frequent, low-risk deployments
  • Supported rapid iteration while maintaining enterprise reliability requirements

technical skills

Languages & Frameworks
Python, TypeScript, JavaScript, FastAPI, Next.js, React, Express.js
Cloud & Infrastructure
Azure (AI Foundry, Azure OpenAI, AI Search, App Service, Functions, CosmosDB, Cache for Redis), CI/CD, Azure DevOps, Azure Container Apps Sandboxes
Databases & Tools
Azure AI Search (vector db), Langfuse (observability), Azure Cosmos DB, Azure PostgreSQL, Azure Cache for Redis, Git, REST APIs, Azure Easy Auth
Specializations
Agent loops and tool calling, RAG, context engineering, MCP servers, sandboxed code execution, LLM observability and evals, multi-tenant LLM platforms
Dev Tools
VS Code, Claude Code, Cursor, Github Copilot CLI, Bash

open source

Autonomous HR Chatbot

442 ⭐ on GitHub

Autonomous HR agent using tools and RAG. Integrates Pinecone, CSVs, Azure Data Lake, and SAP HCM pipelines. Early exploration of agentic workflows (2023) that informed later enterprise production systems.

Python • LangChain • Pinecone • Streamlit • OpenAI API • Azure OpenAI

Model Context Protocol Demo with SSE

Early MCP adoption with SSE and Streamable HTTP. Remote tool integration patterns for Zapier and Gmail.

Python • MCP • FastMCP • SSE • asyncio

AI Assistant for Microsoft 365

LLM assistant integrated with Microsoft 365 (Outlook, Teams, Calendar) to automate tasks and manage communications.

Python • OpenAI Assistants API • Microsoft Graph • Streamlit

Autonomous Mall Assistant

LLM-powered location assistant with fallback recommendations.

Python • LangChain • GPT-4 • pandas • Streamlit

selected writings

Author
Towards AI (1M+ followers)

education

BA, University of the Philippines Diliman, (QS Rank #1, Philippines)

Valedictorian, Bagamanoc Rural Development High School

certifications

  • Microsoft Certified: Azure Data Engineer Associate
  • ITIL 2011 Foundation