0/5 SCORE 0

AI ENGINEER · AGENTIC SYSTEMS · LLM INFRASTRUCTURE

Emanuel Antablin

AI Engineer

I design and ship agentic AI, RAG, and LLM systems that run in production at Walmart Global Tech — across global logistics. The kind of AI that cuts incident response from hours to minutes and saves millions, not the kind that lives in a slide deck.

View the work Get in touch
SCROLL
<0 min Incident MTTR down from ~4 hrs
$0M+ Saved per year workflow automation
0% Throughput lift team velocity
>0% Platform perf global messaging path

How I build

Engineering intelligence, end to end.

Not prompts in a notebook — autonomous systems that observe, reason, and act in production. Four disciplines I bring to every build.

01

Agentic AI

Systems that observe, reason, and act — on their own.

I build autonomous agents that close the loop: they watch live signals, reason over them with an LLM grounded in real context, and take action — paging the right humans with a pre-analyzed diagnosis instead of a bare alert.

  • Observe → reason → act loops (ReAct-style) with tool calling and planning
  • Structured, reliable outputs through deliberate prompt design
  • Confidence scoring + human-in-the-loop override on every decision
  • Deployed across distribution-center server fleets at Walmart scale
02

Retrieval-Augmented Generation

Grounding models in the truth that actually matters.

Agents are only as good as the context they reason over. I wire RAG pipelines that turn operational systems — ServiceNow history, internal APIs, runbooks — into live, queryable knowledge the model can cite at inference time.

  • Chunk → embed → vector search → top-k retrieval → grounded generation
  • ServiceNow tickets + internal APIs as real-time LLM knowledge sources
  • Retrieval tuned against known incidents (cosine threshold, top-k)
  • RAG over fine-tuning where data is dynamic and must stay current
03

LLMs as Infrastructure

The intelligence lives in the pipeline, not the endpoint.

I treat models as components, not products — reasoning engines wired into larger systems through clean interfaces, tool access, and the Model Context Protocol. Model-agnostic by design, swappable as the field moves.

  • Prompt engineering for deterministic, structured machine-readable output
  • Model Context Protocol (MCP) for standardized tool + context access
  • FastAPI services exposing models behind concurrent, async interfaces
  • Provider-agnostic: hosted APIs or self-hosted, chosen per constraint
04

Production & Evaluation

Demos are easy. Production is the job.

A model that works in a notebook is the start, not the finish. I ship intelligence that survives contact with real traffic — measured, observable, containerized, and tuned against the failure modes that actually matter.

  • Precision/recall weighted for false-positive cost in high-stakes domains
  • Per-host threshold tuning to keep agents trustworthy, not noisy
  • Observability with Prometheus / OpenObserve; containerized on Docker + K8s
  • Security baked in — OWASP Top 10 enforced across customer-facing services

Selected work

AI that survives contact with production.

Real systems, real stakes, real numbers — kept at a level I can talk through, with no proprietary internals.

AGENTIC AI · SRE

Walmart Global Tech

Distributed Agentic Monitoring System

Problem

SRE teams discovered incidents reactively — paged with a bare alert, they still had to diagnose from scratch. The discover → escalate → troubleshoot loop ate hours of downtime.

What I built

A fleet of autonomous agents across DC server infrastructure. Each watches its host (memory, CPU, storage), detects anomalies, writes diagnostic notes with an LLM, and pages SRE with a pre-analyzed summary — not just an alert.

Result

Engineers now arrive already knowing what's wrong. Idle discovery time was eliminated and incident response collapsed from a half-day to minutes.

Agentic AILLMPythonKubernetesPrometheus
Read the deep dive
~4 hrs → <30 min
Mean time to resolution
RAG · KNOWLEDGE

Walmart Global Tech

RAG Knowledge Pipeline

Problem

The agents could detect anomalies but not diagnose them — they lacked the historical context, known failure patterns, and runbooks that make an alert actionable.

What I built

A RAG pipeline that ingests ServiceNow ticket history and internal API data, chunks and embeds it, and makes it queryable by agents at inference time — retrieving the relevant past incidents to ground each diagnosis.

Result

Agents surface actionable findings instead of raw alerts, citing the historical signal behind every recommendation. The reasoning layer for the whole monitoring system.

RAGVector SearchFastAPIServiceNowLLM
Read the deep dive
Real-time
Operational context to every agent
APPLIED ML · NLP

Founding Technical Lead · Stealth Startup

PTSD Symptom-Analysis Model

Problem

Clinical PTSD symptom tracking is hard to scale. The goal: detect symptom signals automatically from language patterns in user input — applied ML with real human stakes.

What I built

An end-to-end pipeline I owned from scratch — web-crawler data collection, cleaning and normalization, train/test set preparation, and training of a language-pattern-recognition model.

Result

A working, applied-research ML system where I owned the full stack: the data infrastructure and the model that ran on it, evaluated with false-positives treated as the cost that matters.

NLPSciKit-LearnPythonData Pipeline
Read the deep dive
10×
Relevant sample size, via data pipeline
ML · DATA ENGINEERING

simuwatt

Building Energy-Auditing Model

Problem

A PyTorch model to benchmark buildings and recommend cost-reduction improvements was starved for training data — and the existing data wasn't being fully used.

What I built

The training-data infrastructure behind it: web-scraping pipelines, Python parsers that surfaced overlooked datasets, and Selenium QA automation — plus model-fitting alongside the Data Science lead.

Result

The model shipped and worked: data volume grew sharply and end-of-year user retention rose. A packed early-ML proof point in a four-month tenure.

PyTorchpandasSeleniumPython
All case studies
+20%
User retention · +30% dataset
All case studies, with deep dives →

Toolkit

The stack behind the systems.

AI / ML

  • LLMs
  • RAG
  • Agentic AI
  • Prompt Engineering
  • MCP
  • PyTorch
  • SciKit-Learn
  • TensorFlow
  • pandas

Languages

  • Python
  • TypeScript
  • JavaScript
  • SQL
  • Bash
  • Go
  • C++

Frameworks & Serving

  • FastAPI
  • Flask
  • React
  • PostgreSQL
  • MongoDB
  • Prometheus

Cloud & Infra

  • AWS
  • Azure
  • GCP
  • Kubernetes
  • Docker
  • Jenkins
  • Linux

Trajectory

From ML pipelines to agentic systems.

Lead Software Engineer · Walmart Global Tech 2022 — Present
  • Deployed a distributed agentic monitoring system across DC server infrastructure; autonomous agents detect anomalies, log diagnostics, and page SREs with pre-analyzed findings.
  • Engineered RAG pipelines integrating ServiceNow and internal APIs as live LLM knowledge sources.
  • Led an enterprise agentic-AI initiative across global logistics, reporting architecture and outcomes to director-level stakeholders.
  • Re-architected a legacy messaging platform's recipient resolution as concurrent FastAPI calls — >20% global gain, adopted across all North American distribution centers.
  • Saved the division $1M+/yr by automating technician workflows (+40% throughput); enforced OWASP Top 10 across customer-facing services.
Founding Technical Lead · Stealth Startup 2018 — 2022
  • Trained a language-pattern-recognition model as the core of an AI-led PTSD symptom-analysis pipeline.
  • Built end-to-end data collection, cleaning, and training-set preparation from raw sources.
  • Architected full-stack applications from requirements through deployment; owned the technical roadmap.
  • Implemented OWASP Top 10 countermeasures and penetration testing across servers and web apps.
Software Engineer · simuwatt Aug — Dec 2018
  • Built scraping + parsing pipelines feeding a PyTorch energy-auditing model — expanded the dataset 30% and doubled viable in-house data.
  • Automated QA and internal workflows with Selenium, freeing engineering time for model development.
  • Contributed to model fitting with the Data Science lead — raised end-of-year retention +20%.
Computer Science Tutor · Volunteer 2016 — Present
  • Teach C++ through Python OOP via pair programming, a structured syllabus, and project-based milestones.
B.S. Computer Science — Salutatorian
Full Sail University · Orlando, FL · 2018

Lab

Side experiments.

Mental State Recognition

Applied-research NLP pipeline for PTSD symptom analysis — data collection through model training, owned end to end.

NLPSciKit-LearnData Pipeline

Weather Pattern Analysis

Data-engineering pipeline: external API → ingestion → PostgreSQL → analysis. Clean, reproducible, queryable.

Data EngPostgreSQLPython

Titanic Passenger Risk Prediction

Classic supervised-learning project showcasing feature engineering, model selection, and evaluation discipline.

SciKit-LearnClassification

Contact

Let's build something intelligent.

Hiring, collaborating, or just want to talk agents and RAG? My inbox is open.