input layer
4
Languages
PythonSQLC++Java
The Stack
forward pass, left to right
hidden layer 01
8
Machine Learning
PyTorchTensorFlowScikit-learnDeep Learning
CNNsAutoencodersAnomaly DetectionReinforcement Learning
hidden layer 02
10
Generative AI
LangChainRAGPrompt EngineeringLLM Evaluation
Prompt A/B TestingToken & Cost OptimizationOpenAI API
Anthropic APIGemini APIFAISS
hidden layer 03
8
Backend & MLOps
FastAPIREST APIsWebSocketsJWT Auth
DockerDocker ComposeGitHub ActionsGit
output layer
9
Data & Analytics
PostgreSQLMySQLPandasNumPy
MatplotlibSeabornPower BIEDAStatistical Testing
Projects
three systems, end to end
heart rate74 bpm
modelconv1d-ae
statusnormal sinus
96% accuracy
1D conv autoencoder
WS live inference
- A 1D convolutional autoencoder learns healthy ECG morphology — whatever it reconstructs badly gets flagged.
- FastAPI + WebSockets stream live inference to a React frontend; a Gemini RAG chatbot explains each call in plain language.
PyTorchFastAPIWebSockets
React 18Gemini APIDocker
tokens0
spend$0.0000
cache hits0
2 providers
LRU/TTL caching
$0 CI test cost
- Built from scratch — multi-step pipelines over pluggable OpenAI and Anthropic providers, with per-call token and cost accounting.
- A Mann-Whitney U harness scores prompt variants; a mock provider keeps the whole suite testable at zero API cost.
PythonOpenAI APIAnthropic API
LangChainCaching
PSI0.04
threshold0.20
statestable
KS + PSI tests
Prometheus metrics
1 command deploy
- Drift detection via Kolmogorov–Smirnov and Population Stability Index, alerting to Slack and email on breach.
- Prometheus metrics and a live dashboard, packaged as a one-command Docker Compose stack with Grafana.
PythonPrometheusGrafana
Docker ComposeSlack API
Full write-ups on the work page →
Open to AI/ML & Generative AI engineering roles