AI/ML engineer. I build production LLM systems and the ML underneath them.

Currently forward deployed on enterprise agentic AI. The parts I find most interesting are the ones where an LLM system meets a real constraint, which is usually evaluation, retrieval quality, or inference economics rather than the model itself.

What I work in and what I am interested in. Which of these a given project actually needs varies a lot.

• Agentic systems: LangGraph, LangChain, Google ADK, CrewAI, DSPy, A2A, MCP, Vertex AI Agent Engine, Amazon Bedrock, multi-agent orchestration, function calling, grammar-constrained decoding, JSON-schema tool contracts, ReAct, Plan-and-Execute, graphs as checkpointed state machines, human-in-the-loop interrupts, episodic and semantic memory, maker-checker and critic-verifier loops, self-consistency, log-prob confidence gating, deterministic fallbacks, bounded retries, circuit breakers, idempotent tools, token-exchange identity, sandboxed execution, egress control, prompt-injection testing

• Evaluation: golden datasets, LLM-as-judge, position-bias correction, calibrated rubrics, pairwise comparison, faithfulness and groundedness, bootstrap confidence intervals, recall@k, nDCG, MRR, deterministic replay from recorded fixtures, OpenTelemetry tracing, span-level token accounting, CI regression gates, shadow traffic, canaries, drift monitoring

• Retrieval: hybrid dense and BM25, reciprocal-rank fusion, cross-encoder and ColBERT late-interaction reranking, SPLADE, HyDE, step-back rewriting, query decomposition and routing, contextual chunking, parent-document and sentence-window expansion, metadata filtering, hard-negative mining, contrastive and Matryoshka embeddings, HNSW, IVF-PQ, scalar and binary quantization, incremental indexing, pgvector, FAISS, Qdrant, Elasticsearch, GraphRAG, knowledge graphs in Neo4j, self-correcting text-to-SQL

• Serving: vLLM, TensorRT-LLM, SGLang, continuous batching, paged and radix KV-cache, prefix reuse, chunked prefill, prefill/decode disaggregation, speculative and assisted decoding, FlashAttention, CUDA graphs, torch.compile, FP8 and INT4, KV-cache quantization, custom Triton kernels, model routing, multi-LoRA, SLA-aware scheduling, time-to-first-token, inter-token latency, p99 under load

• Training: FSDP, DeepSpeed ZeRO-3, tensor, pipeline and sequence parallelism, gradient accumulation, activation recomputation, sharded checkpointing, sequence packing, NCCL-tuned collectives, BF16 and FP8, LoRA and QLoRA, SFT, DPO, RLHF, reward modeling, distillation

• Foundations: Python, PyTorch, TensorFlow, Hugging Face Transformers, scikit-learn, XGBoost, tiled CUDA with shared-memory tiling and warp-level primitives, MPI over InfiniBand, ViT, 3D U-Net, Mask R-CNN, CLIP, recommender systems, gradient boosting, probability calibration, conformal prediction, causal inference, lock-free streaming and distributed systems

• Platform: GCP, AWS, Azure, Vertex AI, SageMaker, Azure OpenAI, Docker, Kubernetes, Terraform, CI/CD, Airflow, model monitoring, A/B testing, Databricks, Snowflake, BigQuery, Spark and PySpark, Kafka, dbt, Postgres, MongoDB, Palantir Foundry, Salesforce Agentforce, ServiceNow, Workday, FHIR, NIST CSF, ISO 42001, EU AI Act, HIPAA, GxP

Client work is closed, so these are the public equivalents: reference implementations small enough to clone and run. Each one builds and tests in CI.

• agentic-rag-evals: agentic RAG in LangGraph, gated in CI on eval-metric regression rather than on tests alone.
• gpu-mpi-search: BM25 with cross-encoder rerank, tiled CUDA kernels, data-parallel scoring over MPI.
• mri-segmentation: 3D U-Net for volumetric segmentation, soft-Dice/BCE, patch-based inference.
• warehouse-patterns: dimensional modeling write-up with a runnable DuckDB star schema and SCD-2.
• nutrizy: cross-modal retrieval over a shared text and image embedding space.
• image-captioning-vit: ViT encoder with a Transformer decoder.
• greenbox: CNN disease classifier and LightGBM growth model behind an MQTT edge pipeline.

I write about the engineering at mikitadaroshkin.com, and email is the fastest way to reach me: mikitadaroshkin@gmail.com.