# Samrat Mukherjee — skills

> Technical skills by group, and the four areas the work concentrates in.

## Technical skills

- **AI Agents:** LangGraph, LangChain, agent orchestration, tool / function calling, multi-step workflows, browser-automation agents (computer use), multi-provider routing with fallbacks (Anthropic Claude, OpenAI, Gemini)
- **RAG & Retrieval:** retrieval-augmented generation, embeddings, Pinecone, retrieval-quality evaluation (precision/recall@k, groundedness), grounded answer generation
- **Evals, Safety & Observability:** LLM-as-judge evals, prompt-regression testing, prompt-injection defense, anti-hallucination guardrails, LangSmith, OpenTelemetry
- **Python & APIs:** Python, FastAPI, Pydantic, async concurrency, REST APIs
- **Backend & Distributed Systems:** Go (Golang), Kafka, RabbitMQ, gRPC, PostgreSQL, microservices, event-driven architecture, idempotency
- **Infra & DevOps:** Docker, Kubernetes, GitHub Actions (CI/CD), GCP, AWS

## How he builds agents

### Agent orchestration

Stateful multi-step workflows in LangGraph that plan, call tools, recover on step failure, and escalate to a human when they cannot proceed safely.

Stack: LangGraph, LangChain, tool calling, FastAPI

### Grounded retrieval

RAG over the user's own data so answers come from real background, not from a model improvising. Retrieval quality measured, not assumed.

Stack: Pinecone, embeddings, precision/recall@k, groundedness

### Evals & observability

LLM-as-judge scoring, prompt-regression testing in CI, and every model call, tool invocation and decision traced end to end.

Stack: LangSmith, OpenTelemetry, LLM-as-judge, prompt regression

### Reliability for irreversible actions

Idempotency keys, crash-safe replay and hard anti-hallucination boundaries — because an agent that submits a job application cannot take it back.

Stack: idempotency, crash-safe replay, FactGuard, typed schemas