The Support Agent
RAG-based support that refuses to guess
LangGraph agentCorrective RAGGuardrails
2024 - now
open to agentic-AI rolesBalkrishna Mehta — AI Engineer · Surat, India
I build AI agents that survive production.
Multi-agent workflows, tool calling, RAG over hybrid retrieval, and the operations around them: guardrails, evaluation, latency and cost. I own features from the React screen down to the GPU serving the model. Currently that means Harrir, an AI commerce platform for Iraq.
- search requests a month
- 70K
- search requests a month
- customers served monthly
- 2,300
- customers served monthly
- model versions past the eval gate
- 6
- model versions past the eval gate
- a month for the whole AI stack
- $500
- a month for the whole AI stack
grading production turns with an LLM judgechampion vs challenger before anything shipstool_choice 0.833 → 0.900language accuracy 0.754 → 0.965hot-swapping LoRA adapters, zero restartsone 16GB GPU, chat + vision + embeddingsstatus: deep-work-mode
01 — Now
On the bench this week.
02 — Selected work
The three I'd show first.
03 — Toolbox
The stack under my fingers.
Not a logo wall. This is what's actually in use across Harrir, client work and personal projects: the agent layer, the retrieval layer underneath it, and everything needed to ship and run both.
Agentic
LangGraph
LangChain
Tool calling
Multi-agent
Human-in-the-loop
MCP
Agent evals
Guardrails
RAG
RAG
Hybrid retrieval
Reranking
Embeddings
Qdrant
LoRA pipelines
vLLM
Gemma / TFLite
Product
Python
FastAPI
Next.js
React Native
Flutter
Supabase
AWS
Docker
04 — Writing