AI Engineer at CredResolve, building production AI systems for BFSI collections. Gurugram, India.
I work at the layer where LLMs meet real infrastructure — streaming speech pipelines, agent orchestration under compliance constraints, and control planes where the model proposes and deterministic code decides.
Store Provisioning Platform — Kubernetes-native control plane for multi-tenant e-commerce stores. Level-5 operator maturity: full lifecycle with finalizers, backup/restore, and a closed-loop MAPE-K remediation cycle. The LLM holds zero Kubernetes credentials — it emits typed JSON proposals, and a Kopf operator validates and disposes. Self-hosted vLLM inference, 100% open source, $0 infra budget.
Python Kopf Kubernetes LangGraph vLLM
Inbox Copilot — Grounded email RAG that either cites the exact message or PDF page an answer came from, or abstains. Hybrid lexical + semantic retrieval over a knowledge graph, with a strict citation validator so no claim ships without verbatim evidence. Built against Gmail sync and the Enron corpus.
Python Postgres/pgvector GraphRAG FastAPI
Streaming ASR Service — Production WebSocket STT on IndicConformer. Split gateway (CPU) / worker (GPU) architecture with WebRTC VAD endpointing, Redis token-bucket rate limiting, backpressure, load shedding, and a circuit breaker. Deployed via Docker Compose on EC2.
Python ONNX Redis WebSockets Nginx
Call Center Analysis — Collection-call quality scoring from diarized transcripts. Strict promise-to-pay validation, script adherence, compliance findings cited to turn IDs. One extraction contract, and the model is a fact extractor only — it never decides disposition, escalation, or score.
Python LLM structured extraction
Invoice Resolution Console — PDF invoice → PO match → auditable APPROVE / REVIEW / REJECT. Postgres doubles as database and job queue (FOR UPDATE SKIP LOCKED); LLM handles extraction, Python owns every finance decision. Human review re-runs the same policy engine rather than trusting a free-form verdict.
Python PostgreSQL vLLM FastAPI
- nirmaan-mcp — MCP server + CLI over the Plane API for work-item automation
- WhatsApp Collections Orchestrator — stateful LangGraph campaign flows calling business tools through MCP
- Voice Collections Agent — outbound voice callback workflow with structured outcome logging
Languages Python · TypeScript · C++ · SQL AI/ML LangGraph · vLLM · PyTorch · ONNX · RAG & GraphRAG · MCP · ASR/TTS fine-tuning Infra Kubernetes · Kopf · Docker · PostgreSQL · Redis · FastAPI · Prometheus/Grafana · AWS
The model proposes. Deterministic code disposes.
Every system above draws the same line: LLMs do extraction and intent, never authorization, money, or state transitions. It's why the Kubernetes operator holds the credentials instead of the model, why the invoice policy engine is plain Python, and why the call scorer refuses to let the model pick a disposition.
B.Tech IT, NIT Raipur '26 · GitHub