AI/ML Engineer
NLP, LLM Workflows & Intelligent Automation
AI/ML Engineer with 5+ years in machine learning, NLP and Generative AI. Currently at DICEUS.
Open to roles and collaborations
Most of my work sits between messy data and a decision someone has to trust: cleaning and engineering features, choosing models that can be measured, and checking whether their outputs hold up. I started in 2019 on predictive analytics for business software, moved into credit-risk and fraud work, and now build LLM systems.
My public code on GitHub also includes a Cassandra e-commerce data model, alongside the projects below.
Featured
Fraud detection
Fraud-classification pipeline across 6.3M+ financial records, reaching 95% recall on highly imbalanced data through feature engineering and threshold tuning. Includes a Streamlit app for sub-second risk predictions.
LLM monitoring
A RAG service with a built-in monitoring plane: query-embedding drift detection (PSI, KS-test, centroid cosine) and LLM-as-judge scoring for faithfulness and relevance, with alerts that can trigger re-embedding.
Readmission risk
Predictive analytics dashboard for hospital readmission risk, with patient risk stratification, SHAP explainability and real-time monitoring alerts.
More projects
01Support analytics
Support-ticket analytics platform with AI-assisted ticket categorisation, VADER sentiment scoring, a 7-day volume forecast and an agent performance dashboard.
02Model calibration
Offline toolkit that measures whether classifier probabilities can be trusted (ECE, MCE, Brier score) and repairs them. In its demo, Platt scaling cuts ECE from 0.2219 to 0.0230, a 90% reduction.
03Anomaly detection
Deep-learning anomaly detection for cloud network traffic, trained on CSE-CIC-IDS2018 and MAWIFlow 2025 with attention-enhanced LSTM, parallel CNN and hybrid autoencoder models. Reports 99.98% accuracy.
04Computer vision
Dual-layer verification combining Local Binary Pattern texture analysis with multilingual Hindi and English OCR to detect counterfeit Indian currency. Reaches 95.0% accuracy on a holdout set and runs on standard CPUs.
05Public health analytics
Interactive dashboard exploring 5 years of U.S. Census Bureau SAHIE county-level health insurance data, with a validated ETL pipeline, population-weighted rates, county rankings and CSV export.
Bronze, Silver and Gold lakehouse for e-commerce data with SCD Type 2 tracking, CDC ingestion, a star-schema serving layer and customer segmentation.
07Developer analytics
Ingests repository metadata, commit history and issue trends through REST and GraphQL APIs, with a 2+ year backfill pipeline and PySpark trend analysis.
08Robo-advisory
A robo-advisor that builds risk-profiled portfolios, rebalances when allocations drift more than 5% from target, harvests tax losses and runs 10,000 Monte Carlo retirement simulations.
09Data quality
An active data-quality gate: a dbt and Great Expectations checkpoint that stops the pipeline and sends a webhook alert on schema drift or missing-value spikes, before bad data reaches the marts.
10Retail analytics
Governed retail margin and inventory analytics: a tested dbt star schema feeding a DAX measure library and a Power BI model with row-level security, CI-verified on DuckDB.
Sep 2026 — Present In progress
DICEUS
Built anomaly-detection workflows over 150K+ transaction records, integrated GPT-4 into financial knowledge workflows evaluated against 40+ test queries, and organised data into Bronze, Silver and Gold layers with PySpark and Delta Live Tables. Segmented customers into 5 behaviour groups and cut duplicated implementation effort by 15% with reusable Python components.
Jul 2022 — Aug 2024
Mphasis
Processed 80K+ borrower records for credit-risk classification, prepared 12+ credit attributes and engineered 10+ derived features with Pandas and NumPy. Compared 3 Scikit-learn classification approaches using cross-validation, precision, recall, F1-score and confusion matrices, and served predictions through REST APIs validating 5 key response fields.
Jun 2019 — Jun 2022
Cybage Software
Contributed to a predictive analytics module for business software with Python, Pandas, NumPy and Scikit-learn classification and regression models. Improved dataset completeness by 11% through preprocessing, and improved model performance by 12% through feature engineering and correlation testing.
Aug 2024 — Expected May 2026 In progress
University of Central Missouri
Lee's Summit, Missouri · GPA 3.8 / 4.0
Aug 2015 — May 2019
CMR Engineering College
Hyderabad, Telangana
For roles, collaborations, and AI/ML projects —
saisravan@belvionus.com