generative AI engineering
Gen AI Engineer — Agents, RAG & LLM Systems
I build and ship production LLM systems: multi-agent RAG, vision assistants, NLP-to-SQL and fine-tuned models, with the safety and cost controls they need to stay in production.
Gen AI Engineer at AYE Finance and M.Sc. (Data Science & AI) student at BITS Pilani. 2nd Runner-up, AWS Agentic AI Hackathon.
01
multi-agent RAG · safety · production
Architected and shipped a multi-agent RAG system over credit and policy documents with LangGraph, AWS Bedrock and Qdrant. Layered prompt-injection and output-safety defences (Bedrock Guardrails plus LLM-based validation); delivered through Electron, FastAPI and WhatsApp (Gupshup).
02
citation-grounded RAG · data retention
A NotebookLM-style assistant for employees' own documents: citation-grounded retrieval, so every answer points back to its source, with data-retention controls on uploaded files.
03
OCR · vision · hand-rolled agent loop
Built on ticket data with a framework-free agent loop. Accepts screenshot queries, answers in the user's language, and shares a Tesseract OCR-to-LLM ingestion pipeline reused across RAG and fine-tuning. Cut manual ticket triage by 70% during the pilot.
04
self-hosted · analytics · 5,000+ users
A self-hosted natural-language-to-SQL layer, plus an LLM-generated insight layer, on the analytics platform used by 5,000+ employees.
05
QLoRA · routing · caching
Fine-tuned an open-source LLM with QLoRA for a lending-domain task, and cut production inference cost with tiered model routing and semantic caching (60% cache hit rate).
06
ownership · reliability · team lead
Own the Vault and Repeat services (~40% of company business, ~5 lakh hits a month) plus five other business-critical APIs. Lead 6–7 engineers through sprint planning and code review.
LangGraph, LangChain, hand-rolled agent/job orchestration, RAG, MCP, AWS Bedrock (Guardrails), Groq, NVIDIA NIM, Ollama, prompt engineering and injection defence
RAGAS (faithfulness, context precision/recall), custom cost/latency/cache-hit metrics, LoRA/QLoRA/PEFT, quantisation, semantic caching, tiered model routing
Qdrant, Supabase pgvector, hybrid search and reranking, Tesseract OCR, vision-LLM screenshot queries
XGBoost, PyTorch, Transformers, SHAP, Python, FastAPI, PostgreSQL, Redis, Docker, Kafka, Airflow
SQL, PySpark, Power BI, AWS (Lambda, EKS/ECS, S3, Step Functions, SES, SQS, SNS)
AYE Finance
Build and ship the company's production LLM systems (two multi-agent RAG systems, a multilingual support assistant, NLP-to-SQL, a fine-tuned lending-domain model) and own the Vault and Repeat services plus five other business-critical APIs. Lead 6–7 engineers.
AYE Finance
Led end-to-end design of a production data warehouse and ETL pipelines handling 10M+ records across LOS, LMS and CMS using SQL, PySpark and Python. Mapped complete data lineage, presented the architecture to the CTO and delivered an RBI compliance app 3× faster than planned with a 10-member team. Awarded “Cha Gaye — February” for delivery impact.
RK Agrotech
Built an inventory and lead-management dashboard (HTML/CSS/JS + FastAPI) backed by PostgreSQL, with automated data ingestion from Google Forms and Sheets.
BITS Pilani (Digital)
In progress.
MIT-WPU, Pune
LangGraph · FastAPI · AWS · pgvector · Razorpay · Langfuse
An async, multi-tenant customer-support platform: email and web chat feed a LangGraph supervisor with RAG, vision, action and escalation agents.
LangGraph · RAG · AWS Bedrock · Qdrant · FastAPI · Electron
Multi-agent RAG for credit and policy questions at AYE Finance, with guardrails, private document chat, and desktop and WhatsApp clients.
FastAPI · RAG · Groq · NVIDIA NIM · pgvector · RAGAS
A two-service document platform with RAG chat, YouTube-to-notes and text-to-slides, routing inference across Groq and self-hosted NVIDIA NIM.
AWS · Bedrock AgentCore · Strands Agents · Agentic AI
A workshop entry for the AWS Agentic Football Cup, built with Bedrock AgentCore and the Strands Agents SDK.
XGBoost · SHAP · FastAPI · ML
Prioritises loan applicants with an XGBoost model and explains each ranking with SHAP.