Prijith Ayure — Data & AI Engineer

01

AI Assistant for Data Products

Secure RAG pipeline with PII-redaction guardrails for internal banking docs. Zero data leaves the perimeter.

LangChainRAGBanking

How it works

🔒

PII redaction layer

Before any text reaches the LLM, a custom redaction pass strips account numbers, names, and sensitive identifiers. The model never sees raw customer data.

🗂️

Context-aware retrieval

ChromaDB vector store with metadata filters — queries are scoped by document type and department, so the model retrieves only what's relevant and authorised.

🏦

Zero perimeter exit

Runs fully on-premise. No external API calls, no cloud LLM. Built for a banking environment where data leaving the network is simply not an option.

# Query flow user_query → PIIRedactor.clean(query) → ChromaDB.similarity_search( query, filter={"dept": "risk"}, k=5 ) → LangChain.RetrievalQA.run( context=docs, llm="local-llama" ) → ResponseFilter.validate(answer) → sanitised_response

On-premise · No data leakage · Banking-grade

02

Agentic Research Synthesiser

Privacy-first agent using local embeddings and LlamaIndex. Synthesises papers with zero external API calls.

LlamaIndexLocal LLMsAgentic

How it works

🧠

Local embeddings

All embedding generation runs locally using sentence-transformers. Papers are chunked, embedded, and stored in a local vector index — nothing sent to OpenAI or any external service.

🤖

Agentic reasoning

LlamaIndex agent decomposes complex research queries into sub-questions, retrieves evidence from multiple papers, and synthesises a structured answer with citations.

🛡️

Secure perimeter

Designed for strict data policy environments. The entire pipeline — ingestion, retrieval, generation — runs within a closed network. Zero external API calls by design.

# Agent reasoning loop query = "Compare attention in Llama vs Mistral" agent.decompose(query) → ["How does Llama handle attention?", "How does Mistral handle attention?", "Key differences?"] for sub_q in sub_questions: docs = LocalIndex.query(sub_q, top_k=3) answers.append(LocalLLM.generate(sub_q, docs)) FinalSynthesis.combine(answers) → structured_report

Fully local · Zero API calls · Citation-backed

03

RetireWell — Production API

"We don't ask users to save; we just make their spending work harder for their 60-year-old self." FastAPI retirement engine — SIP recommendations, corpus projection, tax optimisation. Built at BlackRock Hackathon 2026.

FastAPIDockerPythonBlackRock 2026

How it works

📐

Compound modelling

Accurate SIP growth, inflation-adjusted corpus projection, and tax bracket-aware withdrawal planning — not simple interest approximations.

⚡

Async FastAPI

Fully async endpoints handle concurrent requests without blocking. Pydantic schemas enforce strict input validation on every call.

🐳

Docker deployed

Containerised from day one. Single command spin-up, reproducible across environments — built to run anywhere, not just on a dev laptop.

# POST /api/v1/retire/project { "monthly_sip": 25000, "current_age": 28, "retire_age": 55, "expected_return": 12.5, "tax_bracket": "30%" } # Response { "corpus_projected": "₹4.2 Cr", "inflation_adjusted": "₹1.9 Cr", "recommended_sip": 32000 }

BlackRock Hackathon · Production-ready · Not a prototype

View on GitHub →

BuildingIntelligentSystems.

How it works

How it works

How it works

Building
Intelligent
Systems.