LLM & Agent Pipelines
Document AI, OCR and Gemini/GPT structured extraction with consensus checks, RAG and multi-agent workflows that run daily in production.
Jayasurya Sakthivel
Open to full-time roles · MS graduate, May 2026
Data Science & AI/ML Engineer.
I build data systems and AI products that hold up in production — from LLM extraction pipelines and credit-risk models to federated learning research and real-time industrial digital twins.
- 4.00
- GPA, MS @ ASU
- 4
- Hackathon wins
- 1
- Peer-reviewed paper
- 11+
- Shipped projects
Machine Learning✦LLMs✦LangChain✦RAG✦PyTorch✦FastAPI✦MLOps✦PySpark✦BigQuery✦Power BI✦AWS✦Azure✦Databricks✦
About
What I bring to a team.
MS in Data Science, Analytics & Engineering at Arizona State (4.0 GPA). I bridge research-grade modeling with the engineering it takes to ship — pipelines, APIs and dashboards that real teams depend on.
Predictive Modeling
Credit-risk, forecasting and anomaly models — XGBoost, LSTM, SARIMA — with SHAP explanations people can trust.
Data Engineering & BI
ETL/ELT on PySpark, Databricks and BigQuery, star-schema warehouses and Power BI dashboards leadership actually uses.
Applied Research
Federated learning under adversarial corruption and published deep-learning work on underwater image enhancement.
Experience
Roles and impact.
Aug 2026 — Present
AI/ML Engineer
Vizzda
Arizona
3-sourceconsensus extraction2counties automated- Built and deployed end-to-end pipelines for automated real estate document processing, orchestrating Google Document AI, Tesseract OCR and Gemini 2.5 Flash in a 3-source consensus extraction system across Maricopa and Miami-Dade.
- Created entity resolution and skip-trace systems using fuzzy matching, Bizapedia and the Enformion API — with tuned similarity thresholds, cost-control layers and multi-hop LLC traversal — to identify beneficial owners at scale.
- Developed a modular multi-county Python pipeline combining external APIs, Playwright automation, LLM structured extraction and rules-based enrichment to produce contact-ready commercial real estate intelligence daily.
PythonGemini 2.5Document AITesseractPlaywrightEntity ResolutionMay 2026 — Present
Research Assistant
Ira A. Fulton Schools of Engineering, ASU
Tempe, AZ
432experiments run6FL algorithms- Implemented 6 federated learning algorithms combining FedAvg/FedGmd aggregators with SGD, ASS and SASS trainers, running 432 experiments across IID and non-IID MNIST and medical datasets (REFUGE, LAG) with up to 80% corruption.
- Diagnosed why FedGmd hurt accuracy by up to 6% at high corruption — the geometric median collapses toward the corrupt majority past 50% — and confirmed it theoretically via RFA breakdown-point analysis.
PyTorchFederated LearningFedAvgMedical ImagingAug 2025 — May 2026
Instructional Aide & Grader
Arizona State University
Tempe, AZ
100+students supported- Supported 100+ students across Principles of Programming Languages and Operations Research Models through office hours and exam reviews.
- Developed rubrics and solution sets, grading via Gradescope and Canvas to keep academic standards consistent.
Operations ResearchGradescopeCanvasMay 2023 — Jul 2023
Data Analyst Intern
Amrutanjan Health Care Ltd
Chennai, India
300+staff hours saved / yr1M+records processed90%forecast accuracy- Automated ETL and reporting with PostgreSQL, PySpark on Databricks and Python over 1M+ sales records, with anomaly checks flagging data errors before executive reporting.
- Engineered 14 demand and seasonality features with DBSCAN and statistical methods; built LSTM and Seasonal ARIMA forecasts at 90% accuracy, lowering stock-out risk.
- Delivered Power BI variance dashboards that shaped procurement strategy for senior leadership.
PostgreSQLPySparkDatabricksLSTMSARIMAPower BI
M.S. Data Science, Analytics & Engineering
Arizona State University
4.00 / 4.00 · 2024 — 2026
B.Tech Computer Science & Engineering (AI & ML)
Vellore Institute of Technology, Chennai
3.53 / 4.00 · 2020 — 2024
Recommendations
From the people I work with.
“Jayasurya came into Vizzda with little knowledge of commercial real estate, but as a proficient data scientist he had no trouble jumping into things and creating a couple of pipelines for the data we wanted processed efficiently from multiple sources. Any changes/updates I had requested orally or written were swiftly implemented throughout the work process. He would make an incredible asset for any company as an AI/ML Engineer!”
Celso Mundo
Researcher at Vizzda LLC
Managed Jayasurya directly
Projects
Problems, approaches, results.
Case 01
LoanLens
Credit risk decision intelligence · Capstone
1.1M
loan records
Problem — Lenders need default predictions that are accurate on imbalanced data and explainable to borrowers.
Approach — Trained XGBoost, LightGBM and Random Forest on 1.1M LendingClub records with class weighting for a 4:1 imbalance. Added SHAP per-applicant explanations and live FRED macro data at inference.
Result — AUC-ROC 0.703 · 0.67 recall on held-out data, served via FastAPI + React dashboard with a GPT plain-language explainer.
XGBoostSHAPFastAPIReactFRED APIGPT
Case 02
Industrial 3D Digital Twin
1st place · HackAZona, Cognite track
50+
assets modelled
Problem — Plant operators lack a single live view of assets, sensor health and maintenance history.
Approach — Modelled 50+ assets from 2D SVG diagrams into a real-time React Three Fiber scene, wired 20+ sensor streams with anomaly alerts, and added a Gemini/Groq agent with RAG over maintenance logs and SOPs.
Result — Won 1st of 25+ teams. Explainable answers with citations, plus voice-AI callouts to field staff on critical events.
Three.jsR3FNext.jsRAGGroqSupabase
Case 03
Retail Sales Analytics Platform
BigQuery warehouse + Power BI
4 yrs
of sales data
Problem — Four years of retail data with no reliable view of performance, segments or margins.
Approach — Designed a star-schema warehouse in BigQuery with partitioning and clustering, a SQL analytics layer, and DAX-driven Power BI dashboards with forecasting and MoM growth.
Result — Lower scan costs, faster queries, and surfaced high-revenue / low-margin anomalies for KPI tracking.
BigQuerySQLPower BIDAX
Smart ML Framework
Visual ML pipeline builder
Drag-and-drop DAG builder for ML experiments with a FastAPI engine executing blocks in topological order. Dockerized.
TypeScriptReactFastAPIDocker
BusBuddy
Speechbot for blind passengers
Voice assistant using a fine-tuned DistilBERT QA model over transit timetables, fully hands-free with speech recognition and TTS.
PythonDistilBERTgTTS
Voices of History
Talk to historical figures
In-character chat with Llama 3.3-70B plus custom voice cloning (XTTS v2, Bark) behind a voice-first Next.js app.
Next.jsFastAPIGroqXTTS
PathWiseAI
Multi-agent travel planner
Coordinated LLM agents that research, plan and assemble personalized travel itineraries.
Multi-AgentLLMsPython
Skills
Everything orbits Python.
Click any skill group below to put it at the center of the orbit.
SQL
LLMs
Power BI
BigQuery
React
ETL/ELT
Tableau
RAG
ML Modeling
Pipelines
Big Data
Azure
Three.js
PyTorch
Python
Research & Certifications
Published work and credentials.
Publication · 2025
UWE-Net: A Deep Learning Framework for Underwater Image Enhancement Integrating CBAM and Charbonnier Loss
Procedia Computer Science · ScienceDirect
A hybrid CNN combining a VGG16 encoder with a custom decoder and CBAM attention in skip connections, trained with Charbonnier loss for robust underwater image restoration.
Google Advanced Data Analytics
Google · Professional Certificate
Microsoft Certified: Azure Fundamentals
Microsoft
Hackathons
Wins under pressure.
1st
HackAZona v1.0 — Cognite Track
Industrial Digital Twin · out of 25+ teams · 2026
Winner
Principled Innovation Academy Problem Solving Event, ASU
out of 40+ teams · 2025
Best Security
Opportunity Hack
out of 40+ teams · 2025
People's Choice
HackAZona v0.1
out of 30+ teams · 2025
Gallery
The moments behind the resume.
Hackathon stages, demo tables, the paper presentation and the convocation — 10 photos from 2023 to 2026.
Contact
Let's build something that ships.
Open to full-time roles · MS graduate, May 2026. The fastest way to reach me is email.