I build voice agents that hold up in production.

A model-serving platform on Kubernetes (AWS EKS) with a high-accuracy and a fast model variant and an SLA-aware router that shifts traffic to the fast model when the latency SLA is at risk, plus HPA autoscaling, Prometheus/Grafana monitoring and canary rollouts. Under high load, p95 latency fell from 30+ s to 5 s and errors from 5–10% to 0–2%. The Inference Gateway (case study 01) is a from-scratch rebuild of this idea with reproducible measurements.

Full-stack app that turns receipt photos into a per-store price database (LLM vision extraction with structured outputs, human review, rules + fuzzy product matching, unit-price normalization), then solves a MILP to pick which stores to visit, trading item prices against trip costs. Beats single-store and per-item greedy baselines in benchmarks. Also has price-drop alerts, spending insights and a recipe-to-list assistant built on LLM tool calling. Shipped with Docker, GitHub Actions CI, a Helm chart, plan-only Terraform (EKS, RDS, ElastiCache), OpenTelemetry tracing, Prometheus metrics and JSON logs; 158 tests.

Dimensional warehouse built with dbt that runs unchanged on DuckDB locally and on BigQuery (98 of 98 checks on both, the BigQuery build billing 1.29 GiB at $0.008): a star schema with SCD Type 2 customer history, enforced model contracts, PII tagging and masking with role-based analyst views, freshness checks and a per-build data-quality log, all over 1.19M order lines with 84 tests passing in under 5 seconds. Alongside it, a Kafka to PyFlink streaming pipeline computes one-minute order metrics with event-time watermarks and lands them in the same warehouse: 2,000 events/s sustained with results queryable 2.3 s after each window closes, and 18,600 events/s into Kafka in a burst. The same stream also runs on Google Cloud as Pub/Sub → Dataflow (Beam) → BigQuery, measured at the same rate with the latency difference explained. Synthetic data, every number from a real run.

Fraud detection on 284,807 real card transactions (0.17% fraud) with a strict time-based split. XGBoost, Random Forest, LightGBM and logistic regression across three imbalance strategies; unsupervised Isolation Forest and autoencoder detectors; K-means fraud segments (one holds 44% of frauds at 16.6x the base rate); isotonic calibration with a cost-based threshold that cuts expected loss from $2,633 to $646 per 10,000 transactions; SHAP; 5-seed stability. Differentially private training (DP-SGD) with a measured privacy/utility curve; ONNX, int8 and Core ML export at p95 0.04 ms per transaction; Kafka to Spark Structured Streaming scoring at 2,000 events/s that reproduces the batch alerts exactly; PSI drift monitoring. The whole lifecycle also runs as a Vertex AI Pipeline (validate, train, evaluate with a promotion gate, Model Registry, endpoint, batch prediction, weekly schedule): one measured run at test PR-AUC 0.80, endpoint p50 68 ms, about $0.40. The stream also runs on Google Cloud as Pub/Sub → Dataflow (Beam, ONNX Runtime) → BigQuery with per-minute score-drift windows, flagging exactly the same 136 transactions as the batch evaluation.

Retrieval-augmented question answering over 2,067 Wikipedia paragraphs (SQuAD dev) with an evaluation harness that scores answers against gold spans, not just an LLM judge. Compared BM25, dense embeddings (Vertex AI text-embedding-005 in BigQuery vector search) and a hybrid fusion: hybrid retrieves the right paragraph in the top 5 for 94.6% of questions, and Gemini 2.5 Flash reaches EM 0.78 / F1 0.86 with it against 0.09 / 0.14 with no retrieval, at $0.29 per 1,000 questions. Served from FastAPI on Cloud Run; a load test found the async route serialising requests and the fix took throughput from 2.4 to 12 requests/s on one instance. Every number from a real run.

A Java 21 / Spring Boot 3 service that runs YAML evaluation suites against AI agent outputs: strategy-pattern evaluators (contains, regex, JSON Schema, must-not-fire negative assertions, LLM judge), a sealed outcome type, repeat-N banded scoring, per-check precision and recall of alerts, and run-to-run regression diffs, persisted in PostgreSQL. 90 JUnit 5 tests with Mockito and a Testcontainers integration test, 84.7% line coverage. Benchmarked virtual threads against fixed pools on a simulated 50 to 200 ms judge: 7,205 cases/s at a 1,000 bound versus 247 for a 32-thread pool. Every number from a committed results file.

Formulated depth completion as a constrained optimization over the dense depth map (data fidelity, smoothness, edge-aware and coarse-consistency terms, non-negativity by projection). Over 5,000 NYU Depth V2 samples at 5% observed depth, the full objective cut RMSE 33% vs data-only; L-BFGS matched Adam's accuracy at 3.2x the speed. MSML604 team project.

Vapi + n8n voice agent that authenticates callers, retrieves real-time claim status, and logs every call through Google Sheets webhooks, routing callers by claim outcome with stateful tool calls.

Frames grocery shopping as a 0/1 knapsack problem (price = weight, nutrition = value) solved with dynamic programming on Open Food Facts data. Live Streamlit app; code.

Hash-table indexes over a 1.2M-song Spotify dataset for fast artist, song, and album lookup, with a Streamlit search UI.

End-to-end pipeline on FAOSTAT food-balance data (12 countries, 2016–2023): feature engineering, Random Forest / Logistic Regression / XGBoost comparison with stratified CV, and SHAP analysis.

Custom single- and multi-agent warehouse environments with DQN, PPO, and MAPPO policies for placement, picking, and restocking. Reward shaping cut path inefficiency by 18%.

YOLOv5 detection with Kalman filtering and Hungarian matching, reaching 82% MOTA and 76% IDF1 on the MOT dataset.

Hybrid ALS + autoencoder recommender on 15M+ interactions: +12% CTR, +15% watch time, −15% RMSE with Bayesian tuning.

K-means customer segmentation plus a collaborative-filtering recommender for online retail, boosting sales by 20% through personalized suggestions.

Attention-based Seq2Seq Bi-LSTM on CNN/DailyMail, benchmarked against BERT and T5 with ROUGE, BLEU, and perplexity.

OCR + RAG pipeline that extracts structured data from grocery receipts for expense categorization and semantic search.

LSTM forecasting and CNN satellite-image analysis with Kafka streaming and live Streamlit heatmaps for resource allocation.