agentic ai · multi-agent orchestration
Autonomous Customer Support & Action Engine
A production-shaped support platform where agents don't just answer — they route, retrieve, act, verify their own work, and escalate to a human when the stakes are high enough to deserve one.
designed and developed by
Nikhil Chary Sriramoju
GitHub github.com/Nikhil-creat
LinkedIn in.linkedin.com/in/nikhil-chary-sriramoju
Email sriramojunikhil66@gmail.com
Instagram @nikhil__sriramoju
why this isn't just a chatbot wrapper
Every layer earns the word "agentic"
| A2A multi-agent delegation via Agent Cards | RouterSupervisorNode |
| Background agent that self-escalates SLA breaches with no user prompt | ProactiveSLAAgent |
| Self-correction — captures tool tracebacks, repairs its own arguments, retries | ActionExecutorNode |
| Self-verification — a critic agent checks the executor's work before replying | VerificationCriticNode |
| CNN classifies uploaded damage photos, feeds the verdict back into planning | app/vision/ |
| Hybrid retrieval — dense embeddings + BM25, cross-encoder re-ranked | knowledge_base.py |
| Dynamic tool discovery over JSON-RPC, HTTP and stdio transports | mcp/client.py |
| Semantic cache skips the whole graph on near-duplicate questions | semantic_cache.py |
| High-value actions pause for human approval, resume from a checkpoint | interrupt_before |
| PII scrubbed — regex and NER — before every LLM call and every log line | pii_sanitizer.py |
request lifecycle
One message, five agents, one verified answer
01
RouterSupervisorNode
reads Agent Cards, delegates by intent
↓ image attached?
02a
VisualInspectionNode
CNN classifies the damage photo
02b
PolicyRAGNode
hybrid retrieval over support policy
↓
03
ActionExecutorNode
calls MCP tools · retries and self-corrects on failure
↓ high-value or destructive?
—
Human approval queue
graph pauses, resumes from Postgres checkpoint on decision
↓
04
VerificationCriticNode
checks the executor's own output before it reaches the user
running in parallel —
ProactiveSLAAgent polls the approval queue on its own schedule and
escalates anything left pending too long, independent of any chat turn.
built with
Tech stack
orchestration & protocols
LangGraph
AsyncPostgresSaver
Model Context Protocol
A2A Agent Cards
retrieval & vision
BM25
Dense embeddings
Cross-encoder reranking
PyTorch CNN
backend & data
FastAPI
SQLAlchemy 2.0 async
Postgres
Redis
observability & ops
OpenTelemetry
LangSmith
DeepEval
Docker Compose
run it yourself
Up and running in one command
# clone, then:
cd docker
OPENAI_API_KEY=sk-... docker compose up --build
# brings up the API, Postgres, Redis,
# a mock MCP tool server, and an OTEL collector
curl http://localhost:8000/healthz
Full setup, local (non-Docker) instructions, and production hardening notes are in the repository README.