I architect real-time data platforms that turn event streams into decisions — at scale, with sub-second latency.
Data & AI Engineer specializing in streaming lakehouses, agentic AI, and production MLOps. M.S. Computer Science (May 2026, USF) · 3 yrs enterprise data engineering @ Cognizant · AWS Certified Data Engineer
| ⚡ Real-time Kafka → Flink streaming |
🤖 6-agent LangGraph (FinSight) |
📈 600M+ daily records (prod) |
💰 $3.4M+ quantified impact |
🛠️ 6 production-grade builds |
class KoutilyaYenumula:
role = "Data & AI Engineer"
education = "M.S. Computer Science — University of South Florida (May 2026)"
experience = "3 yrs production data systems @ Cognizant"
cert = "AWS Certified Data Engineer — Associate (2025)"
focus = ["real-time streaming", "agentic AI", "medallion lakehouse", "MLOps"]
mission = "Turn raw event streams into decisions before the moment passes."
def ships(self) -> str:
return "production systems, not notebooks-that-die-in-a-drawer"- 🔭 Currently: Shipping real-data analyses (NYC 311, e-commerce conversion) and a HIPAA-aware clinical data platform (FHIR/HL7); building real-time lakehouses with Flink + Iceberg + dbt.
- 🔍 Target Roles: AI Engineer · Data Engineer · ML Engineer (US-based · available now)
- 💡 Passion: Real-time data systems, agentic AI orchestration, production observability, sub-millisecond latency
- 📧 Let's Talk: koutilya718@gmail.com · LinkedIn · Resume
Production-grade systems with real architecture, real tests, and measured impact.
1️⃣ Real-Time Supply-Chain Data Platform — Event-Driven Lakehouse
Problem: Supply chains run on batch data — ERP, WMS, TMS, and IoT signals arrive hours late and siloed, so stockouts and cold-chain excursions are noticed only after the damage is done. Solution: A fully open-source, event-driven lakehouse delivering up-to-the-second visibility across inventory, orders, shipments, and IoT — with streaming ML, RAG assistant, and a what-if digital twin. No managed or paid services.
ERP · WMS · TMS · IoT
│ events
▼
┌─────────────────┐ ┌──────────────┐ ┌──────────────────────────┐
│ Kafka │──►│ Apache Flink │──►│ Apache Iceberg (MinIO) │
│ + Debezium CDC │ │ 4 stream │ │ Bronze → Silver → Gold │
│ + Schema Reg. │ │ jobs · 1min │ └────────────┬─────────────┘
└─────────────────┘ └──────────────┘ │
┌────────────┼──────────────┐
▼ ▼ ▼
┌──────────┐ ┌────────────┐ ┌───────────┐
│ dbt │ │ Trino + │ │ ML / RAG │
│ Gold DQ │ │ FastAPI │ │ + Digital│
│ +lineage │ │ serving │ │ Twin │
└──────────┘ └────────────┘ └───────────┘
Orchestrated by Dagster · Observed by Prometheus · Hardened: PII masking + API auth
| Capability | Detail |
|---|---|
| Streaming core | Kafka + Apache Flink (4 real-time jobs, 1-min windows) |
| Open lakehouse | Apache Iceberg on MinIO — Bronze → Silver → Gold |
| CDC + governance | Debezium CDC · Avro/Schema Registry · PII masking · API auth |
| Serving & ops | Trino + FastAPI · Dagster–dbt lineage · Prometheus alerts |
Kafka · Flink · Iceberg · Trino · dbt · Dagster · Debezium · MLflow · MinIO · FastAPI
2️⃣ ReadmitGuard — 30-Day Readmission Prevention (Healthcare) · FHIR/HL7 · HIPAA · Iceberg
🔗 github.com/koutilyaY/readmitguard
Problem: Medicare penalizes hospitals up to 3% of reimbursement for excess 30-day readmissions (HRRP) — but the signal is buried in siloed, standards-heavy clinical data. Solution: A real-time clinical platform that ingests HL7 v2 ADT + FHIR R4, lands it in a HIPAA-aware medallion lakehouse (Safe Harbor de-identification), and scores readmission risk at discharge — returning a FHIR RiskAssessment.
| Aspect | Detail |
|---|---|
| Standards | FHIR R4 · HL7 v2 ADT · ICD-10 / LOINC / RxNorm |
| Clinical methods | LACE index · Charlson comorbidity · CMS-HRRP cohorts |
| Compliance | HIPAA Safe Harbor de-id · MRN tokenization · audit log |
| Stack | Kafka → Spark → Iceberg → dbt → Dagster · XGBoost · FastAPI |
FHIR · HL7 · Kafka · Spark · Iceberg · dbt · Dagster · XGBoost · FastAPI
3️⃣ FinSight AI — Multi-Agent Financial Intelligence · LangGraph · Kafka · Spark
🔗 github.com/koutilyaY/finsight-ai
Problem: Investors manually analyze market data for hours — slow, error-prone, expensive. Solution: End-to-end MLOps platform with 6 ReAct agents (LangGraph) over a Kafka → PySpark medallion → Delta Lake pipeline, with Feast/Redis features and streaming FastAPI/Streamlit serving.
| Aspect | Detail |
|---|---|
| Agents | 6 ReAct agents orchestrated with LangGraph |
| Pipeline | Kafka/Avro → Spark medallion → Delta Lake → Feast/Redis |
| ML | XGBoost + Isolation Forest anomaly detection · FinBERT sentiment |
| Serving & Ops | FastAPI (async SSE) · Streamlit · Prometheus/Grafana/Jaeger · Terraform/EKS |
Kafka · PySpark · Delta Lake · LangGraph · XGBoost · SHAP · MLflow · FastAPI (async SSE) · Redis · Prometheus/Grafana/Jaeger · Terraform EKS
4️⃣ PayGuard — Real-Time Fraud Detection · Kafka · Spark · LightGBM
🔗 github.com/koutilyaY/payguard-realtime-fraud
Problem: Card fraud happens in milliseconds; batch detection is 24 hours too late. Solution: Live Kafka events → Spark Structured Streaming → Delta Lake medallion → LightGBM scoring with real-time decisions and an analyst feedback loop.
| Aspect | Detail |
|---|---|
| Streaming | Kafka → Spark Structured Streaming → Delta medallion |
| Model | LightGBM scorer (synthetic data, MLflow-tracked) |
| Signals | velocity · high-value · account-takeover |
| Serving | FastAPI decision API · Streamlit dashboard |
Kafka · PySpark Structured Streaming · Delta Lake · LightGBM · MLflow · FastAPI · Streamlit · Prometheus/Grafana
→ Live Dashboard
5️⃣ DataShield — Real-Time Data Observability · FastAPI · Anomaly ML · Lineage
🔗 github.com/koutilyaY/DataShield
Problem: Pipeline failures cascade silently; heavyweight tools are slow. Solution: Lightweight FastAPI service for real-time schema validation, anomaly detection, and lineage/blast-radius tracking — with a zero-infra demo mode.
| Aspect | Detail |
|---|---|
| Tests | 22 / 38 pass with no infra |
| Detection | anomaly ML on statistical baselines (synthetic) |
| Validation | real-time schema checks + lineage / blast-radius |
Python · FastAPI · PostgreSQL · Pydantic · Docker · Prometheus · GitHub Actions
6️⃣ DocuSense — Contract Intelligence Agent · Local-first RAG · Ollama · FAISS
🔗 github.com/koutilyaY/docusense
Problem: Legal teams spend days reviewing contracts for obligations, exceptions, risks. Solution: Local-first RAG (FAISS + Ollama) over legal contracts — cited Q&A plus a risk-classification agent that flags obligations & risk clauses. 100% local, privacy-first.
LangChain · RAG · FAISS · Ollama · FastAPI · Streamlit
7️⃣ CreatorPulse — Creator-Economy Burnout Prediction · Causal Inference · Survival Analysis
🔗 github.com/koutilyaY/creatorpulse
Problem: Creator churn quietly erodes platform revenue, but standard models can't separate correlation from causation or predict when burnout hits. Solution: End-to-end DS system on synthetic behavioral data combining XGBoost prediction with Causal Inference (DiD), Survival Analysis, and A/B testing to quantify retention levers.
XGBoost · Causal Inference (DiD) · Survival Analysis · A/B Testing · Python
Public, real-world datasets — the messy-data rigor and honest reporting behind the platforms above.
📊 NYC 311 Response-Time Analysis — 3.46M real requests · DuckDB · dbt · SQL
🔗 github.com/koutilyaY/nyc-311-analysis
Every NYC 311 request filed in 2024 (3.46M rows, live NYC Open Data) analyzed for how fast the city actually responds. Shows the citywide "average" is a statistical trap — NYPD's ~1-hour closes hide housing's 4-day and taxi's 74-day tails — normalizes complaints per capita, and measures the heat-vs-temperature effect (r = −0.80) instead of asserting it. dbt medallion, data contracts, tests, and a right-censoring fix that changed the headline numbers.
DuckDB · dbt · SQL · Dagster · Streamlit · multi-source join · right-censoring
🧪 E-Commerce Conversion — is the model worth shipping? · XGBoost · SHAP · MLflow
🔗 github.com/koutilyaY/online-shopper-conversion
Real UCI session data (12,330 sessions) predicting purchase intent — validated forward in time, not with a flattering random split. Catches a near-leakage feature, benchmarks the model against a one-line business rule, and honestly recommends against shipping when the rule wins. Pandera contracts, MLflow tracking, FastAPI serving, model card.
XGBoost · SHAP · MLflow · FastAPI · pandera · temporal validation · PR-AUC
Migrated 11 legacy ETL jobs to Databricks medallion; cut costs 47%.
| Achievement | Result |
|---|---|
| Monthly savings | $22K/month (47% reduction) |
| Pipeline uptime | 87% → 98% |
| Incident response | 4 hrs → 35 min |
| Prevented loss | $1.8M (data quality framework) |
| Daily records processed | 600M+ (Airflow + S3 + RDS) |
Languages
Streaming & Processing
Orchestration & Cloud
AI / ML & Serving
Observability
Where I go deep
Streaming & Real-Time Kafka · Flink · Spark ▰▰▰▰▰▰▰▰▰▱
Lakehouse & Modeling Iceberg · Delta · dbt ▰▰▰▰▰▰▰▰▰▱
Cloud & Infra AWS · Docker · K8s · TF ▰▰▰▰▰▰▰▰▱▱
ML / MLOps XGBoost · MLflow · SHAP ▰▰▰▰▰▰▰▰▱▱
Agentic AI / LLM LangGraph · RAG ▰▰▰▰▰▰▰▱▱▱
| 🛰️ Streaming-first | Data has a shelf life — decide before the moment passes. |
| 🧱 Medallion or it didn't happen | Bronze → Silver → Gold, with contracts at every hop. |
| ✅ Tests are infra | Schema contracts, dbt tests, freshness SLAs — not optional. |
| 🤖 Agents that act | Pipelines that detect, decide, and remediate themselves. |
| 📟 Observable by default | Prometheus + lineage > hoping nothing broke overnight. |
