Tsion Ebuy
Tsion Ebuy

Building AI that works with real data.

AI engineer in Los Angeles. Agents, RAG and data pipelines.

See my work Contact
Portrait of Tsion Ebuy
120/120agent tests passed
0.90model test accuracy
41,638rows cleaned
1.00chatbot correctness

Work

Multi-agent · Udacity capstone

NovaMart support agents

An orchestrator sends each request to the right specialist agent.

  • 4 agents on Bedrock AgentCore
  • 3 knowledge bases, searched in parallel
  • Guardrail tested on the deployed runtime
  • 120 of 120 automated tests passed
AgentCoreStrandsRAGDynamoDBX-Ray

REQUEST FLOW

AVERAGE LATENCY PER NODE

Orchestrator30.7 s
PolicyAgent18.7 s
RefundAgent9.3 s
InventoryAgent8.4 s
CommAgent7.7 s
KB lookups~0.6 s
Terminal output showing the observability checks passing and a final score of 120 out of 120 points.
Test run: 120/120.
X-Ray trace map of one request across the orchestrator, agents and knowledge bases.
X-Ray trace map
X-Ray node list with latency and request rates for each agent and knowledge base.
Node latency list
Three real test conversations

I want to return my order ORD-29001, it is broken.

Return approved. Premium window is 60 days. Refund $1,099.00.

I want to return my order ORD-28001, it is the wrong size.

Standard window is 30 days, and it closed on October 1, 2026.

Is Amazon cheaper than NovaMart?

Blocked by the guardrail. I can help with orders, returns and NovaMart policies.

Routing · Udacity

Support chatbot flow

Classifies each message and routes it to the right path.

  • Classifier prompt at temperature 0
  • Bedrock Evaluation: 1.00 correctness
Bedrock FlowsLambdaDynamoDB

ROUTING

message classifier temperature 0 bug report platform question other request
Data warehouse · University of Calabria

Traffic accidents warehouse

A star schema for nationwide accident data.

  • One fact table with severity
  • Date, location and weather dimensions
  • Designed from ER model to PostgreSQL
Star schemaDimensional modelingPostgreSQL
Star schema in pgAdmin: an accident fact table with severity, linked to date, location and weather dimension tables.
Star schema in PostgreSQL
Dimensional fact model with weather, location and date hierarchies.
Dimensional fact model
Entity-relationship diagram with accident, date, weather, location and road entities.
ER design
Machine learning · University of Calabria

Heart disease classifier

Sorts public CMS indicator records into six conditions.

  • 42,640 rows and 29 columns cleaned to 41,638 and 9
  • SMOTE for class balance, 71 features
  • MLP scored 0.90 on a held-out test set
Pythonscikit-learnSMOTE

ACCURACY BY MODEL

MLP (test)0.90
Random forest0.89
C4.5 tree0.87
CART tree0.87
Naive Bayes0.75
k-NN0.53

Other models on a validation split.

Confusion matrix of the MLP on the test set for the six conditions.
Confusion matrix, test set

Experience

GW AppsFeb 2024 to now

Software Developer, AI Integrations

LLM features for business software: record summaries, text categorization and document data extraction. Prompt design and output testing.

Handshake AIMay 2025 to now

AI Trainer and Evaluator

Rubric-based audits of LLM outputs, Elo ranking and benchmark task design. Freelance.

UCLA2024

Research Intern, Agricultural Water Management using AI

Sole data analyst. Python and Pandas cleaning, Power BI and Matplotlib dashboards.

Skills

AI and agents

Amazon BedrockAgentCoreStrands SDKBedrock FlowsPrompt designEvaluation

RAG and safety

Knowledge BasesS3 VectorsTitan embeddingsGuardrails

Cloud

LambdaDynamoDBS3IAMCloudFormationCloudWatchX-RayCI/CD

Data and code

PythonPandasSQLPostgreSQLMySQLMongoDBSnowflakePower BITableauJavaJavaScript

MSc AI and Computer Science, University of Calabria (in progress)

Future AWS Agent Engineer, Udacity Nanodegree

BSc Computer Science and Engineering, Mekelle University

Generative AI with LLMs, DeepLearning.AI and AWS

Data Science and AI certificate, UCLA

Advanced Data Analytics, Google

Let's talk.

Tsiongher2013@gmail.com LinkedIn