Applied AI & backend engineer

I build LLM systems that read messy documents,
find the evidence, and act on it.

Open to AI and backend engineering internships and roles.

Final-year CSE (AI & ML) student at VIT Bhopal. I spent eight months shipping production AI at Ywork.ai in London (document pipelines, RAG and agents) and I'm now building Clarivo, an agentic accounts-payable system.

01

Clarivo

Autonomous accounts payable: three-way matching and exception resolution with LLM agents.

  • Gmail Pub/Sub ingestion, containerized Lambda extraction, hybrid retrieval, then a cited and grounded LLM verdict per invoice.
  • Grounding layer checks every cited figure against its source; unverified citations are flagged, not trusted.
  • Event-driven re-evaluation: a new delivery re-opens closed invoices so split-shipment overbills resolve themselves.
Django REST FrameworkAWS LambdaDocker · ECRS3DynamoDBWeaviateGeminiTesseract OCRGmail APIGCP Pub/Sub
02

HindSite

Semantic search over the pages you actually read, with a two-stage retrieval pipeline.

  • Engagement-based capture (reading time, scroll depth) indexes only pages actually read.
  • Two-stage retrieval: pgvector similarity over 1024-dimensional Cohere embeddings, then cross-encoder reranking.
  • A query router separates semantic search from live tab-switching, by voice or text.
FastAPIPostgreSQLpgvectorCohere embeddingsCross-encoder rerankingChrome extension

Ywork.ai

AI Engineer Intern

Remote, London, UK

May 2025 – January 2026

Production AI features at a London construction-procurement startup.

Letter of recommendation (opens in a new tab)
  • Built document-intelligence pipelines on AWS Lambda combining OCR with LLM-based extraction, converting unstructured construction PDFs (quotes, invoices, spec sheets) into schema-validated JSON for downstream services.
  • Built a tool-calling folder-management AI agent that reorganizes uploads into a clean hierarchy: embedding files into Weaviate, retrieving related files by similarity, and having an LLM group them by content, filenames and structure; new synced files are placed automatically, and chat corrections are applied via tool calls, retaining context.
  • Engineered the RAG layer over a multi-tenant Weaviate corpus (chunking, embedding and semantic retrieval) powering a real-time AI chat over client project data.
  • Integrated the Google Maps API to surface nearby suppliers for a location, letting users shortlist local vendors and request quotations in-platform.
  • Built event-driven file-ingestion REST APIs in Django (webhook and polling pipelines over AWS S3) keeping cloud files synchronized, and architected OAuth 2.0 authorization in Django REST Framework across 4 cloud providers (Microsoft, Google Drive, OneDrive, Dropbox), handling token refresh and per-provider scope differences.

I'm a final-year student who spent eight months building production AI features at a London construction-procurement startup. I care about the parts of LLM systems that decide whether they can be trusted: retrieval quality, grounding, evaluation and failure modes, and about backends that stay boring and reliable.

VIT Bhopal University

B.Tech, Computer Science and Engineering (AI & ML)

2023 – 2027

CGPA 8.75 / 10

LLMs and retrieval

LLM agentsTool callingStructured outputsPrompt engineeringRAGHybrid searchSemantic searchVector embeddingsCross-encoder rerankingLLM evalsDocument intelligence and OCR

Backend

PythonDjango and DRFFastAPIREST API designEvent-driven architectureWebhooksOAuth 2.0Multi-tenant systemsWebSockets

Data

PostgreSQLpgvectorDynamoDBWeaviateChromaDBSQL

Cloud and tooling

AWS (Lambda, S3, DynamoDB, ECR)GCP Pub/SubDockerGit

Models and frameworks

LangChainGeminiOpenAICohere

Let's talk about what you're building.