Résumé
Open to graduate roles · London or relocating anywhere in the UK · No visa sponsorship needed until late 2028
Hello there

Hi, I’m Nikhil Rana.

I build AI that shows its working.

I’m an MSc Artificial Intelligence graduate working across responsible AI, machine learning, data science and data engineering. I built an LLM system that answers questions about UK company accounts straight from their own machine-readable tags, which lifted accuracy from 45.1% to 92.2%.

Try the live app See the research Download CV
Tailor this page · I’m interested in…

Pick a role and the page will reshape itself: a tailored summary here, and the relevant skills, projects and experience highlighted as you scroll.

◭Fig. 01 — The Himalaya, drawn as a ridgeline chart. This is the view from where I grew up, rendered as data. Move your cursor and the range shifts in depth. Click on empty space to send a ripple, and turn the sound on in the bottom-left corner for a singing bowl.

LLM accuracy 92.2%, up 47.1 points. 7,106 facts parsed. 60 of 60 filings parsed. Detector caught 0 of 3 errors. 900,000+ ISP accounts segmented. MSc on track for a Distinction. No sponsorship needed until late 2028. Willing to relocate UK-wide.

01 — About

I’m the quiet one who reads the footnotes. I care about the people behind the data, and an answer that only sounds right isn’t good enough for them. It has to be provably right.

I grew up in Nepal and studied Computer Science & IT at Tribhuvan University, where databases, statistics and simulation became my favourite tools. The course took me from Database Management Systems through Data Warehousing & Data Mining to Advanced Database and a Database Administration elective, and my first data role was at WorldLink Communications, Nepal’s largest internet provider.

In 2025 I moved to London for an MSc in Artificial Intelligence at the University of West London, where I’m on track for a Distinction. For my dissertation I engineered a full pipeline, from raw filings to parser, database, retrieval, LLM and evaluation, to find out whether AI can be trusted with company accounts.

I’m open to roles in responsible AI, ML engineering, data science, data engineering, databases and analytics. What drives me is simple: technology should be honest with the people who rely on it. I like to understand a problem deeply before I solve it, I listen more than I talk, and I do my best work on things that matter to someone.

What I bring

Responsible AI, measured

I build evaluations with gold answers, bootstrap confidence intervals and honest error reporting, not just demos.

Data engineering & databases

I built an iXBRL parser and a SQLite fact store that parsed 60 of 60 filings with zero failures, grounded in DBMS and SQL coursework.

Machine learning & LLMs

My work spans RAG, LLM evaluation across Claude and GPT, clustering, regression, LightGBM and deep learning coursework.

Clear communication

I’ve written reports and presented findings directly to managers, and I work to compliance-critical procedures every week.

0%
LLM accuracy on UK company filings, up from 45.1%
0
Tagged facts engineered into a structured database
0
Customer accounts at the ISP where I built segmentation models
Distinction
Expected MSc classification, with a 70%+ average

How I build AI for people

5 principles · tap to expand
Location · Relocation

Based in London. Happy to move.

I’m willing to relocate anywhere in the UK, and I’d especially welcome roles in Leeds, Newcastle and across the North. Hover over or tap a city.

Selected: London (home base).
02 — Experience

Where I’ve worked

From analysing 900,000+ customer accounts to keeping a busy London store compliant and stocked.

MAR 2024 — JUN 2024
Lalitpur, Nepal · Full-time placement
Nepal’s largest ISP: 900,000+ consumer accounts across 73 of 77 districts. Mentored by the Marketing Manager.

Data Analyst Intern

WorldLink Communications Ltd · Marketing Department
  • Customer segmentation: built a K-Means pipeline in Python on tenure, set-top boxes, usage charge, bandwidth and plan days remaining. I standardised the features, chose k with the Elbow Method and visualised the clusters with PCA.
  • Churn analysis: built a three-cluster churn-risk model to flag high-risk segments for retention campaigns.
  • Data preparation & reporting: collected and cleaned data in Excel and Python, analysed a 1,600-respondent, seven-region survey in SPSS, and presented the findings to the team.
  • Competitive mapping: researched rival ISPs’ offers and wrote a positioning report for the Marketing Manager.
  • Across the team: analysed campaign performance with the Digital Marketing Manager, wrote an app-monetisation report for myWorldLink, and provided data support for a dual-band router promotion.
  • Engineering discipline: wrote unit and system tests covering loading, preprocessing, scaling and clustering.
Segment 0Tight, homogeneous group suited to niche offers
Segment 1Varied attributes suited to a broad approach
Segment 2Small niche suited to highly personalised offers
Segment 3Largest and most diverse, needing versatile strategies
PythonScikit-learnK-MeansPCASPSSExcelPowerPoint

Tap any skill to see where I’ve used it.

OCT 2025 — PRESENT
London · Part-time

Trading Assistant

Sainsbury’s
  • Compliance: I work across frozen, fresh, produce and meat departments following strict food-safety and time-temperature controls, alongside a full-time MSc.
  • Operations: I process warehouse deliveries, fulfil Uber Eats and Deliveroo orders, and use barcode-driven stock systems to prioritise replenishment.
Stock systemsComplianceCustomer serviceTime management
03 — Featured research · MSc dissertation 2026

An AI that reads company accounts by their tags

Every number in a UK company filing already carries a machine-readable label. I built an LLM system that uses it.

UK accounts are filed in iXBRL (Inline XBRL). The document looks like a normal web page to a person, but every figure is wrapped in a tag that says exactly what it is: its accounting concept, period, currency, scale and any breakdown. My system parses those tags into a database and answers questions only from the tagged facts, and if the fact isn’t there, it says so.

Tested on 12 filings (Tesco, Lloyds, AstraZeneca, National Grid and more) · 102-question benchmark · Claude Sonnet 4.6 & GPT-5.6 Terra · Supervised by Dr Ali Gheitasy

The app is live. Upload any UK Companies House iXBRL filing (up to 25 MB) and ask questions in plain English. Every answer is traced to the tagged fact behind it. 10 free questions per session on GPT-5.6 Terra; Claude Sonnet 4.6 by access code. The assistant below replays results from my study.Free hosting, so the first load can take up to a minute to wake up.Open the app
0%

Accuracy with tag-aware retrieval

Up from 45.1% (+47.1 pp, 95% CI [37.3, 56.9]). GPT-5.6 Terra replicated the gain: 43.1% → 84.3%.

52% vs 2.9%

The AI didn’t lie. It went quiet.

Without retrieval, the models mostly said “not found”, and confident wrong answers were rare.

0 / 3

The detector was blind

A SelfCheckGPT-style check caught none of the errors, because financial facts can have more than one valid answer.

1 · Choose a filing
2 · Choose the system
3 · Ask a question
iXBRL assistant · pick a filing and a questionready
04 — Projects & hackathons

Things I’ve built

All code on GitHub
01
Responsible AI

Ethical Framework for NLP Summarisation

An evaluation framework for Meta AI’s BART dialogue-summarisation model covering bias mitigation, privacy and output transparency. It asks whether a model should be deployed as well as whether it works.

NLPBARTAI Ethics
02
Data science

Discount Impact & Customer Segmentation

Do discounts really drive bigger spends? I segmented customers with K-Means, predicted high-value transactions with Logistic Regression and Decision Trees, and measured discount elasticity with Linear Regression (MAE, R²).

Scikit-learnRegressionClustering
03
Computer vision · Security

Facial Recognition Voting System

A secure academic voting system that verifies voters with LBPH and Haar Cascade facial recognition, then confirms them with one-time-password two-factor authentication.

OpenCVPython2FA
04
Machine learning · Healthcare · Team of 4

Diabetes Prediction: fewer missed diagnoses

An MSc Machine Learning group project on around 800 patient records. We replaced medically impossible zeros, engineered features and compared six models across four resampling strategies, picking winners by the fewest missed diagnoses rather than accuracy alone. A tuned voting ensemble with an optimised threshold reached 84.4% accuracy with 0 false negatives on the 77-patient test set.

XGBoostLightGBMSMOTE-ENNEnsembles
01

FRC Hackathon

Financial Reporting Council

Investigated LLM hallucination failure modes on structured UK iXBRL filings and restructured XBRL schemas for efficient AI ingestion, working hands-on with Companies House data. This fed directly into my dissertation.

Regulator-led
02

VLGE AI Hackathon

AI experiences

Pitched a technical concept for immersive, AI-driven experiences and placed third.

3rd Place
03

Exoplanet ML Hackathon

Scientific machine learning

Built a probabilistic forecasting pipeline with LightGBM quantile regression on HDF5 data, modelling uncertainty as well as point estimates.

Uncertainty ML
05 — Tools & Domains

What I use, and where I apply it

Tools

Bars show how much I’ve used each tool, from coursework through to daily use.

Domains

Responsible AI & governance

Hallucination evaluation, bias, privacy and transparency, UK GDPR, and ethics-approved research.

Evidence · MSc module · Dissertation · BART framework

Data engineering & databases

Parsing messy structured data, schema design, SQL and SQLite, and data warehousing and mining.

Evidence · Dissertation pipeline · DBMS, Data Warehousing, Advanced Database & Database Administration modules

ML & LLM systems

RAG, LLM evaluation, clustering, regression, gradient boosting and deep learning.

Evidence · Dissertation · Hackathons · MSc modules

Financial reporting & RegTech

iXBRL, IFRS vs FRS 102, extension concepts, and Companies House and FCA sources.

Evidence · Dissertation · FRC Hackathon

Customer analytics

Segmentation, churn risk, survey analysis and competitive mapping for a 900k+ customer ISP.

Evidence · WorldLink internship

Retail operations

Stock systems, deliveries, online fulfilment and food-safety compliance.

Evidence · Sainsbury’s
06 — Education

Where I trained

Explore both degrees semester by semester. Anything highlighted maps directly to data and AI roles.

Sep 2025 — Sep 2026

MSc Artificial Intelligence

University of West London, UK
DistinctionExpected classification · 70%+ average
2020 — 2024

BSc Computer Science & IT

Tribhuvan University, Nepal
2:172.43% · UK Upper Second Class equivalent

Modules from the TU BSc CSIT curriculum, including my Database Administration elective. Highlighted modules map to data and AI roles.

07 — Beyond the desk

FPL, football & everything else

Quiet in the room, but I notice a lot. I collect data even when nobody asks me to, and on weekends it all goes into Fantasy Premier League.

My FPL season in data · 2026/27

Back in form this season.

Live from the official FPL data · Gameweek 5 · I’ve played every season since 2017/18.

360Total points
387,467Overall rank · top 4%
85Best GW · GW1, rank 37,833 (top 1%)
£100.6mTeam value
Points per gameweek
Overall rank (higher is better)
🏆 Best finish: 266,096 (top 3%, 2021/22) 📈 Five straight seasons in the top 7% worldwide (2020/21–2024/25) 🧤 59 points left on my bench so far. As a goalkeeper, I know how the bench feels.
The FPL lab · captaincy model

Who gets the armband?

Choose a player type or move the sliders. My model weighs each factor and gives a verdict. Switch to Gut brain to see how Saturday-morning me decides.

0
Captaincy score
A toy model for fun, not financial or FPL advice. Every point costs.
Nikhil smiling on a sunny clifftop above a pier and the sea
Travel

Chasing natural places

Coastlines, hills and anywhere with a horizon. Growing up near the Himalaya will do that to you.

Between the sticks

5-a-side goalkeeper

I’m the last line of defence, which is good practice for catching problems before they hit the back of the net, on the pitch or in a data pipeline. Clean sheets are not guaranteed, but I always dive in.

Live

Concerts

Nothing beats live music: the crowd, the energy and a song you’ll never hear the same way twice. It’s also why the page chimes when you click the mountains with the sound on.

☠︎
Favourite anime

One Piece

A story about loyalty, friendship and chasing a dream across an impossibly big world. I’m still on the voyage.

✎
Writing

On Medium

I write to think clearly, and explaining things simply is a skill worth practising. Read on Medium ↗

Over a cup of tea, you’d learn that…
I’d rather have one deep conversation than ten small ones.
NepalBorn & raised
BSc CSITTribhuvan University
WorldLinkFirst data role
UKMSc AI · London → wherever’s next
08 — Contact

Let’s build AI
people can trust.

I enjoy getting to know people, so say hello even if you’re not hiring. I’m based in London, happy to relocate anywhere in the UK, and I reply to every message.

LinkedIn /in/nikhilrana2076 GitHub @NikhilRana2076 Medium @nikhilrana2076 Résumé PDF

Available from autumn 2026 · Graduate visa, no sponsorship needed until late 2028.