Open to internshipsfull-time from Feb 2027London, UK
I spent three years keeping enterprise IT running. Now I'm doing an MSc in AI and building systems I can show you the numbers for.
01About
For three years I was the person on the other end of the ticket: triaging fifty incidents a day, administering Active Directory and keeping SLAs green for enterprise clients. Near the end I built a Python tool that used a language model to turn prose access requests into concrete directory changes, and I approved every one. That tool is where AI stopped being a subject and became the work.
The model did the interpretation, which is what it is good at. I did the authorisation, which is what I was accountable for. It saved me two to five hours a week, and it taught me that the interesting problem is the system around the model.
02Work
One live system I keep improving, one tool that ran in production, and five data projects, each written up with what went wrong as well as what worked.
fig. 2 is an illustrative run. The other figures are drawn from each project’s own results.
03Assistant
I built an AI assistant that answers questions about my work from my own case studies, decision records and code. Every answer shows the sources it came from, so you can check it rather than trust it.
01
Your question
in plain English
02
Hybrid search
keyword (BM25) and semantic
03
Fuse and rerank
RRF, then an LLM grades each passage
04
Grounded answer
written only from those passages, with sources
04Experience
2021 to now · 4 roles
01 / 05Timeline
to· 10 mos
Tech MahindraFull-time
Remote, India
to· 11 mos
Tech MahindraIBU TACFull-time
Noida, India · On-site
to· 2 yrs 7 mos
CapgeminiFull-time
Noida, India · On-site
The technical bridge between front-line operations and development teams, across enterprise environments.
toNow
Northeastern University LondonPart-time
London, UK · On-site
Represent Northeastern University London at open days and offer-holder events.
Next
Open to AI and ML internships, including a summer 2027 internship or a placement that forms an assessed part of my MSc, and to full-time roles from February 2027, once the MSc finishes in January.
London · UK-wide · Remote (UK)
05Skills
I rate every skill against evidence you can check. Familiar means coursework or tutorials. Working means I’ve shipped something real with it. Strong means production use, on my own. Open any row to see the evidence.
Showing 17 of 17 skills: 2 strong,10 working,5 familiar
06 skills · 4 working · 2 familiar
Building now, as the subject of my MSc dissertation (agentic AI and LLMs) and in the portfolio assistant.
Working: Shipped something real
A working retrieval pipeline: BM25 + semantic search fused with RRF, an LLM reranker, grounded generation.
Working: Shipped something real
Streaming Gemini answers through my own provider layer with a fallback model; an LLM tool at Capgemini that drafted directory changes for review.
Working: Shipped something real
Regression on 175,777 rows (R² 0.8800 on all features), and fraud and retention classifiers: logistic regression, a tuned random forest, RFE and RFECV.
Working: Shipped something real
IIIT Bangalore data science programme with an NLP specialisation, awarded Oct 2025.
Familiar: Coursework or tutorials only
Coursework and tutorials only. No trained model in a public repo yet.
Familiar: Coursework or tutorials only
03 skills · 3 working
An LLM tool used daily at Capgemini, plus the fraud and retention models and analyses of 175,777 delivery orders and 3,014,248 taxi trips.
Working: Shipped something real
Cleaning and feature engineering, including cutting a 3,014,248-row dataset to 489,542 usable rows.
Working: Shipped something real
MSc Database Management Systems: ER model to a 3NF schema, eight queries and tested indexes. Then 29 MySQL questions on crop insurance data, with CTEs and window functions.
Working: Shipped something real
05 skills · 2 working · 3 familiar
The portfolio assistant is a TypeScript codebase; MSc web coursework.
Working: Shipped something real
Daily use across coursework, the analysis projects and the assistant.
Working: Shipped something real
Tutorials and coursework only.
Familiar: Coursework or tutorials only
AWS Cloud Technical Essentials certificate; no production use yet.
Familiar: Coursework or tutorials only
BCA and diploma coursework.
Familiar: Coursework or tutorials only
03 skills · 2 strong · 1 working
50+ incidents a day against SLA for two and a half years at Capgemini.
Strong: Production use, independently
Accounts, access permissions and endpoint compliance across enterprise environments.
Strong: Production use, independently
Case management for Netgear and Orbi support at Tech Mahindra.
Working: Shipped something real
06Education
04 programmes · 04 certificates
to
Northeastern University London
Programme
Machine learning, deep learning, statistics and applied AI, with research-oriented coursework and peer learning groups.
Modules · 08
to
International Institute of Information Technology Bangalore, with upGrad
Four graded courses: Python, data tooling, machine learning and advanced machine learning for NLP, assessed through assignments, projects and exams.
Verify certificate(opens in a new tab)Result
3.85
CGPA out of 4.0
to
Chandigarh University
The fundamentals of computer science and application development, studied alongside full-time work.
to
Kalinga University, Raipur
Three-year diploma. Final project: a Unity 3D program that presents 3D images on a solid surface.
Result
81%
First division
Each opens its public verification page
07Contact
It’s in London
Availability
Open to internships · full-time from Feb 2027
Open to AI and ML internships, including a summer 2027 internship or a placement that forms an assessed part of my MSc, and to full-time roles from February 2027, once the MSc finishes in January.
Flagship · live 2026
A retrieval-grounded assistant that answers recruiter questions about my work, with every claim traced to a source document. Hybrid BM25 + semantic search fused with reciprocal rank fusion, an LLM reranker, and answers written only from what was retrieved, checked against an 86-question golden set. It is live: try it below.
An illustrative run of the pipeline, not real output. The question “What has he built with LLMs?” is searched two ways, with BM25 and with vector search. The two rankings are fused with reciprocal rank fusion, an LLM reranker grades each passage from 0 to 3, and the answer cites the two passages that answer it: capgemini-automation.md and core-questions.md.
Illustrative answer: At Capgemini he built a Python tool that used an LLM to turn prose access requests into proposed Active Directory changes. He approved every one himself.
Access requests arrive as prose
The model interprets.I authorise.The model has no direct write.
Production tool · Capgemini 2025
A Python tool that read prose access-request tickets, used the company's internal model to propose a concrete Active Directory change, and put it in front of me to approve. The model interpreted; I authorised. It never touched the directory on its own.
Retired when I left: it had one owner and no handover
A unit chart of the 300 held-out validation claims, one dot each: 74 fraudulent and 226 honest, the natural fraud rate of 24.7%. The dots are ordered by outcome, not placed by any feature: in each block the claims the model flagged come first, from the top left. Showing Logistic regression. Logistic regression, with a cutoff of 0.60, caught 62 of the 74 fraudulent claims and missed 12; it sent 34 honest claims for review anyway and cleared 192. F1 0.729, AUC 0.855; accuracy 0.875 in training and 0.847 held out, a gap of 2.8 points. The tuned random forest caught 54 and missed 20; it flagged 36 honest claims and cleared 190. F1 0.659, AUC 0.825; accuracy 0.982 in training and 0.813 held out, a gap of 16.9 points. The forest overfits: before tuning it scored 1.000 on every training metric. Logistic regression beats it on every validation metric.
ML · classification 2025
Fraud detection on 1,000 auto insurance claims. A six-feature logistic regression caught 84% of the held-out frauds at 85% specificity and beat a tuned random forest on every validation metric, even though the forest scored 1.000 on its training data before tuning. Its strongest feature is a label leak in the synthetic data, which the write-up points out.
Odds ratios from the final retention model, for eight of its fifteen features, each against its baseline:
Senior level: 12.05 times the odds of staying. Remote work: 5.62 times the odds of staying. Mid level: 2.64 times the odds of staying. Male: 1.80 times the odds of staying (a protected characteristic). Overtime: 0.72 times the odds of staying. Work-life fair: 0.34 times the odds of staying. Work-life poor: 0.29 times the odds of staying. Single: 0.18 times the odds of staying (a protected characteristic). Values above 1 mean more likely to stay than the baseline; below 1, less likely.
ML · classification 2025
Logistic regression on 74,610 employee records, built so HR can see what it reacts to. On 21,168 held-out employees it recognised 74% of those who stayed and 74% of those who left, within half a point of its training scores. Senior staff have 12 times the odds of staying of entry-level staff. Gender and marital status predict too, and the write-up says HR should never use them to treat an individual differently.
Data · SQL 2025
Twenty-nine questions in MySQL on PMFBY, India's crop insurance scheme, over 1,870 district-year records, using joins, subqueries, CTEs, window functions, keys and constraints. Farmers paid 13.1% of the Rs 43,007 crore gross premium, and their share fell every year, from 19.1% in 2018 to 10.6% in 2021, while the premium nearly doubled.
The query reads FarmersInsuranceData and, for each srcYear, sums GrossPremiumAmountToBePaid and divides it by 100 to give gross_crore, the gross premium in rupees crore, then takes 100 times the summed FarmersPremiumAmount divided by the summed gross premium to give farmers_pct, the farmers' share. It groups and orders by year and returns 4 rows. 2018: Rs 7,263 crore gross; farmers paid Rs 1,390 crore (19.1%), state governments Rs 2,987 crore (41.1%), the central government Rs 2,886 crore (39.7%). 2019: Rs 9,214 crore gross; farmers paid Rs 1,279 crore (13.9%), state governments Rs 4,186 crore (45.4%), the central government Rs 3,749 crore (40.7%). 2020: Rs 13,411 crore gross; farmers paid Rs 1,560 crore (11.6%), state governments Rs 6,413 crore (47.8%), the central government Rs 5,439 crore (40.6%). 2021: Rs 13,119 crore gross; farmers paid Rs 1,395 crore (10.6%), state governments Rs 6,155 crore (46.9%), the central government Rs 5,570 crore (42.5%). Over 2018 to 2021: Rs 43,007 crore gross; farmers Rs 5,624 crore (13.1%), states Rs 19,740 crore (45.9%), the centre Rs 17,643 crore (41.0%). The gross premium nearly doubled while the farmers' own payment stayed between Rs 1,279 and Rs 1,560 crore, so their share fell every year, from 19.1% to 10.6%.
Test R² and mean absolute error on the held-out set, for each feature set:
ML · regression 2025
Linear regression over 175,777 Porter orders. All 12 engineered features scored test R² 0.8800; Recursive Feature Elimination down to 8 cost six points of R², the opposite of what the brief expected and the most useful finding.
Tip percentages are highest between 10pm and 2am. A single month of 2023 records is 3,014,248 trips; the working sample is 489,542 trips drawn from all twelve months. 3 figures in the submitted notebook were corrected.
Data · exploratory analysis 2025
EDA over 489,542 trips sampled across all twelve months of 2023 NYC taxi data. Tipping peaks between 10pm and 2am, and card payments tip far more than cash. Three figures in the submitted notebook turned out to be plotted from placeholder data, so I published the corrections alongside it.
Also
Notes appCoursework · MSc Web Services · 2026
A notes application built for the Web Services module of the MSc. University coursework rather than production work.
