Open-source contributions, personal projects, and works in progress across MLOps, AI/ML, DevOps, and Cloud Engineering.
Contributed to the Go-based local LLM serving engine. Helped improve REST API documentation, multi-model management, and integration examples for production pipelines.
Contributing to the production-grade LLM serving engine. Focus on improving Kubernetes deployment guides, multi-GPU configuration docs, and OpenAI-compatible API examples.
Contributed integration examples for LLM tracing and evaluation. Added Python notebooks showcasing MLflow model registry workflows with Kubernetes-based model deployment patterns.
Contributing pipeline component documentation and example DAGs. Focus on Katib hyperparameter tuning configurations and multi-cloud deployment guides for ML workloads.
Added community integrations and example chains for production RAG systems. Contributed MLOps-focused tutorials showing how to pair LangChain with MLflow for LLM observability.
Contributing to Ray Serve deployment patterns and Ray Data pipeline examples. Focused on Kubernetes-based Ray cluster setup guides for large-scale ML training workloads.
Contributing DAG templates for ML data preprocessing pipelines. Added Kubernetes Pod Operator examples and MLflow integration patterns for production data workflows.
Added example flows for ML data ingestion pipelines. Contributed Docker and Kubernetes deployment recipes and async task coordination patterns for real-time data processing.
Contributing fine-tuning recipes and deployment scripts. Focused on production-serving patterns with vLLM/BentoML integration and low-rank adaptation (LoRA) training examples.
Contributing MLOps integration guides connecting Dify's agentic workflow engine with production observability tools. Added examples for Kubernetes deployment and scaling.
Always looking for contributors and co-maintainers for MLOps tooling, LLM deployment patterns, and cloud infrastructure automation. Let's build together.
Every agent tool call passes a deny-by-default policy, rate limiter and session budget, and lands in a hash-chained audit log that can prove nobody rewrote history. Path traversal caught by canonicalization. Pure stdlib, 22 tests, CI green.
BM25, character-trigram vectors and reciprocal rank fusion from first principles. Failure modes isolated per query flavor: BM25 collapses on typos, fusion recovers everything BM25 loses. Backed by 19 tests, CI green.
A numpy MLP with manual backprop reproduces catastrophic forgetting: naive fine-tuning drops task A from 99% to 56%. L2 anchoring and replay measured against it. Backed by 10 tests, CI green.
Measures how chunking strategy changes RAG retrieval quality. BM25 built from scratch, golden-set evaluation, and the boundary-loss effect pinned as a test. Backed by 16 tests, CI green.
The MCP tools subset from first principles: JSON-RPC server and client with schema validation and correct error codes, in about 200 readable lines. Backed by 14 tests, CI green.
LoRA implemented from first principles in NumPy. Exact merging, frozen-base guarantees and rank trade-offs, every claim pinned as a test. Backed by 9 tests, CI green.
Exact and IVF vector search from first principles, with the recall-versus-probes trade-off every vector database sells, measured on lumpy data. Backed by 14 tests, CI green.
Reproducible training with metric-regression gates on AUC and recall, leak-free evaluation and generated model cards. A retrain that loses quality fails CI. Backed by 15 tests, CI green.
PSI and KS drift detection as a pipeline gate. Baseline-quantile bins so skewed features cannot hide drift, unseen categories treated as maximum signal. Backed by 14 tests, CI green.
A policy gate over terraform plan that blocks destroys and replaces of protected resources in CI, before they reach production. Backed by 13 tests, CI green.
Error-budget arithmetic and generated multiwindow burn-rate Prometheus alert rules, following the Google SRE Workbook policy. Backed by 16 tests, CI green.
Discrete-event simulation showing why continuous batching beats naive LLM serving: 6x throughput and 9x lower p95 latency on the demo workload. Backed by 9 tests, CI green.
One API in front of many LLM providers: cheapest-first routing, response caching, automatic failover with per-provider circuit breaking, and a cost ledger separating spend from spend avoided. 19 tests, runs offline in one command.
Golden-set evaluation for RAG systems as a deterministic CI gate: a change that degrades answer quality fails the build and names the broken case. Baseline comparison, hard must-include gates, retrieval scored separately from generation. 29 tests.
Reads NVIDIA DCGM exporter output and re-exports burn, waste (the idle share of spend, in dollars) and cost per 1k inferences as Prometheus metrics. Strict parsing, flagged fallback pricing, 503 over stale data, stdlib only. 15 tests.
A production-ready MLOps template with MLflow experiment tracking, DVC data versioning, GitHub Actions CI/CD, Docker containerization, and Kubernetes deployment manifests.
End-to-end RAG application with FastAPI backend, LangChain orchestration, Pinecone vector store, and React frontend. Deployed on AWS ECS with auto-scaling and CloudWatch monitoring.
Complete Infrastructure as Code (IaC) for spinning up an ML platform on AWS — EKS cluster, SageMaker endpoints, S3 data lake, RDS feature store, and CloudWatch dashboards.
Built a feature store using Feast + Redis + Kafka for real-time ML feature serving. Includes Airflow DAGs for batch feature computation and a FastAPI endpoint for online serving.
A Dockerized observability stack for LLM applications — MLflow for experiment tracking, Prometheus + Grafana for metrics, Langfuse for LLM traces, and alerting via PagerDuty.
Multi-agent framework using LangGraph + Ollama for autonomous task decomposition and execution. Supports tool use, memory, and human-in-the-loop approval workflows.
Delta Lake + BigQuery + Dataflow pipeline on GCP for real-time and batch ML feature engineering. Includes dbt transformations, data quality checks, and Vertex AI integration.
Automated model retraining pipeline triggered by data drift detection. Uses Evidently AI for drift monitoring, GitHub Actions for retraining, and automatic promotion via A/B shadow testing.
AWS Lambda + API Gateway facade for multi-provider LLM routing (OpenAI, Anthropic, Ollama). Includes cost tracking, rate limiting, prompt caching, and fallback routing logic.
Hybrid classical-quantum machine learning experiments built on IBM's Qiskit SDK. Implements variational quantum circuits (VQC) for classification tasks, quantum kernel methods, and QAOA for combinatorial optimization. Runs on IBM Quantum real hardware via IBM Cloud and local Aer simulator.
End-to-end AI agent automation pipelines built with n8n and Anthropic's Claude API. Workflows include: multi-step web research agents, RAG retrieval pipelines, automated email drafting from Slack triggers, GitHub PR summarisation bots, and LLM-powered data enrichment flows. All orchestrated visually in n8n's node graph with zero backend boilerplate.