luv. Let's Talk

Open to 2027 graduate roles

Hi, I'm Luv

An AI and data engineer building retrieval systems that show their sources.

Open graphite mechanical keyboard case with a screen in the lid and an orange hinge
LUV AGRAWAL / CREATIVE INDEXTYPING

Retrieval you can check. Pipelines that ship.

  • 2 internships
  • ✦
  • 94/100 GraphRAG accuracy at k=5
  • ✦
  • 100,500 records in under 2 seconds
  • ✦
  • 90%+ degree average
  • ✦

Numbers first, adjectives never.

0query accuracy gain, GraphRAG over a hybrid baseline (54/100 to 94/100 at k=5)
0graph nodes and 57,050 edges across 1,000+ NHS documents
0synthetic patient records processed in under 2 seconds on AWS Lambda
0UK Tier 1 bank as first client of the contract intelligence platform

Demo

See it work

Forty-six seconds from a stack of PDFs to an answer with receipts.

A 46-second illustrative walkthrough of the contract intelligence flow, built from sample data. It is not a recording of any client's product, and the evaluation figures are from my I/O Atelier GraphRAG work.

Try it

Ask a sample contract

Every answer shows its receipts. Try to catch it guessing.

Sample lease · illustrative text

4.2Rent is payable quarterly in advance on the usual quarter days.
9.1The tenant shall insure the premises against the risks listed in Schedule 3.
12.3Either party may end this lease on six months' written notice given on or after the third anniversary of the term commencement date, provided all rent due has been paid.
15.4The service charge is capped at the rate stated in Schedule 5 and reviewed annually.
18.1Any tender bond is released on practical completion.
Pick a question to see a cited answer.

Pre-written sample answers over made-up text, to show the idea. Nothing here is a live model or client data.

Experience

Two internships, shipped in parallel

Two teams, one keyboard, no spare weekends.

HaloRFP · July to September 2026

From blank page to a bank as first client

AI and Data Engineering Intern. I built a contract intelligence platform for commercial real estate and construction, from initial design to first client deployment.

  • PDF and DOCX ingestion with clause-aware chunking of legal text
  • LLM extraction into a JSON schema of obligations, break clauses, payment terms and risk flags
  • Fine-tuned a small Qwen3-4B language model
  • Every answer cited back to its exact source passage
  • Full technical handover so the team could keep building
Ingestion · live pass1 / 3
PDFTender pack
ObligationsBreak clausesPayment termsRisk flags
Cited to source passage
I/O Atelier · June to September 2026

GraphRAG that beat the baseline by 74%

AI and Data Engineering Intern. I built a GraphRAG layer over a BM25 and vector Hybrid RAG baseline and evaluated it on 500 clinical queries across five top-k settings.

  • Knowledge graph of 23,576 nodes and 57,050 edges persisted in pgvector
  • Retrieval confidence up 27 points, 81% vs 54%
  • Removed three classes of false positives before scaling to 1,000+ documents
  • Event-driven AWS ingestion with least-privilege IAM and Neo4j loading
Evaluation · k=5500 queries
0graph nodes
0graph edges
0confidence
University of West London · 2024 to 2027

BSc Computer Science, 90%+ average

Final year. Strong theory and mathematics foundations, which is where the retrieval and evaluation instincts come from.

  • BMW Group UK Data Analytics Final Assessment Centre, December 2025
  • Mathematics Olympiad: Distinction (Grades 5 and 7)
  • Microsoft x EDI Diversity in Tech Competition, 2025
  • AWS Cloud Practitioner Foundations: in progress
Key resultsFinal year

The graph

A knowledge graph you can walk

A document is a list. Meaning is a map.

Dense knowledge graph visualisation: thousands of green nodes joined by grey edges, with a few large dark hub nodes at the bottom right
Knowledge graph visualisation. Each dot is an entity pulled from the documents, each line a relationship.

Plain vector search finds passages that look similar. It misses answers that live across several documents. The graph connects the entities those documents share, so a query can walk from one document to the next.

  • Entities extracted with DeepSeek-V3 and stored in pgvector
  • NetworkX traversal across 8 to 14 documents per query
  • Scored on 500 clinical queries across five top-k settings
  • Exposed to AI assistants as MCP tools, with no direct database access

Fine-tuning

Teaching a small model to do one job well

Bigger is easy. Right-sized is the craft.

  1. 01

    Task

    Pick one narrow job: structured extraction from contract clauses

  2. 02

    Data

    Build and clean a labelled set of clauses paired with target JSON

  3. 03

    Fine-tune

    Adapt Qwen3-4B so it follows the schema instead of free-writing

  4. 04

    Evaluate

    Compare against the base model on held-out clauses before trusting it

  5. 05

    Ship

    Serve the small model for the extraction step in the pipeline

↻ A 4B model is cheaper and faster to run than a large general model, so the fine-tune is judged on accuracy and cost together.

Cloud and data

Built on AWS, split across shards

When one database carries everything, one bad day carries everything.

Live sketch: EC2, sharded Postgres on RDS, Terraform

 

EC2

Application servers behind a load balancer. New instances join when traffic grows, and requests go to the least busy one.

RDS PostgreSQL

Managed Postgres, one database per shard, so backups, patching and failover are handled by AWS.

Sharding

Each record is routed by a hash of its key. No single database carries all the data or all the writes.

Terraform

The whole stack is declared as code, so the same environment can be created again with one command.

Illustrative diagram of the pattern, drawn with simulated traffic. It is not a live system or a client environment.

How I build

From raw documents to answers you can check

Boring pipelines make exciting answers.

01 Ingest

Parse everything, chunk by meaning

Contracts, leases and tender packs arrive as PDF and DOCX. I parse them, split legal text clause by clause rather than by character count, and keep parties, key dates and document type as metadata.

02 Extract

Turn text into a schema

An LLM converts clauses into a defined JSON schema: obligations, break clauses, payment terms and risk flags, ready for downstream legal review.

03 Retrieve

Hybrid search plus a graph

BM25 and vector search find candidates, then a knowledge graph traverses across 8 to 14 documents per query. Keyword-overlap thresholds, a confidence floor and citation recalibration cut the false positives.

04 Prove and hand over

Measure it, then document it

Every change is scored on hundreds of real queries. When the work ships, the team gets the architecture, runbook, troubleshooting guide, API reference, decisions log and cost model.

Ingestion · clause chunksmetadata
Clause
Clause
Clause
PartiesKey datesDocument type
Extraction schemaexcerpt
{
  "obligations": [ ... ],
  "break_clauses": [ ... ],
  "payment_terms": { ... },
  "risk_flags": [ ... ],
  "source_passage": "exact text"
}
Retrieval · graph traversal8 to 14 docs
BM25VectorGraph
Evaluation · k=5500 queries
ArchitectureRunbookTroubleshootingAPI referenceDecisions logCost model

Toolkit

What I build with

Sharp tools, kept few, used daily.

AI and RAG

  • GraphRAG
  • Hybrid RAG
  • BM25
  • Vector search
  • Knowledge graphs
  • NetworkX
  • Entity extraction
  • RAG evaluation
  • LLM fine-tuning
  • Qwen3
  • DeepSeek-V3
  • MCP
  • LangChain
  • Prompt engineering

Programming

  • Python
  • SQL
  • FastAPI
  • REST APIs
  • Pandas
  • NumPy
  • Boto3
  • GeoPandas
  • Matplotlib
  • Seaborn

Cloud

  • AWS
  • Lambda
  • S3
  • IAM
  • CloudWatch
  • EC2
  • RDS PostgreSQL
  • Terraform
  • Serverless
  • Event-driven pipelines
  • Docker

Data and tools

  • PostgreSQL
  • Database sharding
  • Supabase
  • pgvector
  • Neo4j
  • Git and GitHub
  • Jupyter
  • ClickUp
  • Agile Scrum
  • Technical documentation

Good systems start with hello.

Have a role, a project in mind, or just want to say hi? I'm looking for a 2027 graduate role in AI or Data Engineering.

luvga07@gmail.com

GitHubLinkedInDownload CV

Hand-built in London, on a keyboard that doubles as the menu.

© 2026 Luv Agrawal · London, UK · Back to top