I work at the intersection of business and information systems. I use data and technology to answer practical business questions: what's happening, why, and what to do next.
- 🎓 Master of Business IT, La Trobe University, Australia (2025–2027): business analysis, data and systems thinking, digital transformation
- 🎓 BE Computer Science (AI & ML), Ramaiah Institute of Technology, Bangalore (2021–2025), CGPA 8.5
- 💼 Marketing Coordinator, SKBC Realty (2025): digital campaigns, lead generation, client relations
- 📝 Author of several research publications; I also do project coordination, mentoring and volunteering
| Area | Tools & skills |
|---|---|
| Data processing & analysis | Python, pandas, NumPy, data cleaning, EDA, forecasting & backtesting, data structures & algorithms, Jupyter |
| SQL & analytics | SQL, SQLite / DuckDB, dbt (star schemas, tested models), KPIs, cohorts, funnels |
| Applied AI | LLM APIs, prompt design, text-to-SQL with guard rails, LLM evaluation, text classification & summarisation |
| APIs & automation | REST APIs, ETL pipelines, data-quality testing, GitHub Actions |
| Cloud basics | Cloud storage, serverless functions |
| BI & reporting | Power BI (data modelling, DAX, dashboards), Streamlit dashboards, Excel, Matplotlib / Plotly |
A one-page tour of these projects, each with its headline finding and key chart, is on my portfolio site (source: built from one data file and deployed with GitHub Actions).
| Project | What it does |
|---|---|
| Skills-Demand Dashboard | Which occupations and regions are hiring in Australia, from Jobs and Skills Australia's monthly count of online job ads (Internet Vacancy Index, August 2026 release), with a Streamlit dashboard and every report rerun by GitHub Actions on every push. Read naively, the files show an 18% "slump" from October to January every year and rank tiny cells as the fastest growers; with those traps removed, ads are −3.3% on the year but +21.2% on 2019, and demand has moved to regional Australia: regional areas are +67.1% on 2019 against +4.9% for the capitals, led by doctors and nurses. Ends in 6 ranked recommendations for recruiters, training providers, employers and analysts |
| AusTender Procurement Spend | Spend, competition and compliance analytics in DuckDB SQL and Python on 59,464 Australian Government contract notices from the official AusTender API (FY2024-25), rerun by GitHub Actions on every push. A dashboard built straight from the feed would overstate spend 2.6 times; after fixing that, only $4.17bn sits in categories concentrated beyond one contract, while $19.11bn of common services is spread across 20+ agencies and long supplier tails. Limited tenders almost always cite an allowed reason, and half of late notices come in catch-up batches. A training-spend categoriser (keyword rules plus a Claude backend) is scored against 288 hand-labelled notices. Ends in 8 ranked recommendations for Finance, procurement leads and internal audit |
| Purchase-to-Pay Process Mining | Process mining in DuckDB SQL and Python on a real purchase-to-pay event log (BPI Challenge 2019, 1.6M events): 21 compliance rules, where items wait, and what work is redone or worth automating, rerun by GitHub Actions on every push. It finds the controls held (10 of 21 rules had no failures) and that only 5.0% of the mean 83.0 days from PO to first clearing is process overrun; the rest is mostly agreed terms and the payment-run calendar. The rework that costs time is a PO price changed after the invoice: 1,860 items holding 14.6% of value, 92.2% blocked and 4.0 days slower to clear. Ends in 8 ranked recommendations for procurement, accounts payable, finance, audit and treasury |
| NEM Demand Forecasting | Explainable half-hourly electricity demand forecasts for Australia's 5 NEM regions on AEMO data: a rolling backtest, a blend of smoothing and a corrected 3-year average, archived weather forecasts and tested MW buffers, rerun by GitHub Actions on every push. On a held-out year it reaches 3.3-8.5% day-ahead error, 22-45% below each region's best simple baseline. It finds that a "95%" buffer built the usual way covers only 66% of days, and that a fixed timetable timed a battery as well or better in 4 of 5 regions. Ends in 8 ranked recommendations for energy retail and battery teams |
| Airline On-Time Warehouse | A tested star-schema warehouse (dbt-core on DuckDB) for Australian domestic airline punctuality since 2004, from government open data, plus a Power BI model with 31 DAX measures and check queries, rebuilt and tested by GitHub Actions on every push. The raw file stacks four grains, so a naive sum overstates flights 3.51×. It finds punctuality still below pre-COVID (77.9% on time vs 79.4% in FY2018-19), and that on busy routes the airline matters more than the route: on Brisbane–Melbourne, Qantas ran 88.4% on time and Jetstar 59.3%. Ends in 6 ranked recommendations for a corporate travel team and its data team |
| Victorian Rental Market Analysis | Where rents are rising fastest across Victoria, and who can still afford to rent, from 25 years of Homes Victoria rental-bond data joined to ABS inflation, building approvals and Census figures (DuckDB SQL, an audit that reconciles every parsed table with the published report). It finds the biggest rises were in regional cities, not Melbourne: rents there rose 46.9% since 2019 and the share of new leases a lower-income household can afford fell from 46.8% to 36.4%. Ends in 5 ranked recommendations for councils, planners, agencies, employers and investors |
| LLM SQL Report Helper | Plain-English business questions → guarded SQL → a short report whose numbers are checked against the result. Personal data and writes are blocked inside SQLite, every metric has one written definition, and any backend (Claude, or an offline semantic layer) is scored on 32 gold-answer questions plus a reworded held-out set. It shows why that matters: the offline path is right 59% of the time on the wording it was built from but 19% on reworded questions |
| Data-Quality Pipeline on DuckDB | dbt-style data tests written from scratch (YAML checks compiled to SQL), run on Melbourne's live pedestrian-sensor feed, plus a trust scorecard for 7 business reports and a daily GitHub Actions gate that goes red only on new errors. It finds 3 of the 7 reports shouldn't be used as usually built, and that year-on-year growth is +6.2% like-for-like, not the +3.8% the published totals show |
| Marketing Lead Funnel Analysis | DuckDB + SQL analysis of 8,000 marketplace seller leads joined to ~99k orders: channel and landing-page conversion, a shift-share split of a monthly drop, and whether signed sellers start selling. It finds only 43% of signed sellers made a sale within 90 days, and that paid search beats organic search once selling is counted (55% vs 39% start selling) |
| Data Structures in Practice | Hash maps, heaps, a sweep line and union-find applied to 342k NYC 311 service requests, each benchmarked against naive code and a library call. It finds 1 in 6 tickets repeats an issue that's already open, and that matching nearby reports by location catches nearly 3× as many of the water department's own duplicates than exact addresses do |
| Serverless Property Data Lake | AWS S3 + Lambda pattern for HM Land Registry's monthly sold-price file: validated, address-free Parquet, SQL in place, a least-privilege SAM template and a cost estimate (~$0.05/month). Run offline on an S3 emulator. It shows 96% of new-build sales are registered 7+ months late, so the latest months understate the market |
| LLM Review Categoriser | Sorts ~1,000 product reviews into business themes (Claude backend plus an offline keyword baseline), checks accuracy against a hand-labelled sample, and splits complaints into fixable sub-issues. It finds 1 in 11 negative reviews describes a product that doesn't work, a problem a theme view alone hides |
| Retail KPI Dashboard | Interactive Streamlit dashboard on 1M retail transactions: KPI tiles with year-on-year deltas, country/product drivers and a customer revenue bridge. It shows a flat +2.3% year hiding heavy churn, with £1.1M of lapsed customers to win back |
| E-commerce Funnel & Delivery Analysis | DuckDB + SQL case study on ~99k marketplace orders: fulfilment funnel, lost revenue, late delivery vs reviews, retention cohorts and seller vs carrier delays. It finds 72% of late orders left the seller on time, pointing to destination-based delivery promises |
| FX Rates Pipeline | A self-updating ETL (public API → validated SQLite, run daily by GitHub Actions) plus an SQL report on currency risk for an Australian importer. It shows that re-forecasting the budget rate quarterly cuts the average monthly cost miss by 43% |
| Retail Data Cleaning & EDA | Cleans 540k messy retail transactions with a logged, rule-by-rule pipeline, then finds seasonality, customer concentration and lapsed-customer revenue, ending in 7 ranked recommendations |
| SQL Business Analytics | 8 business questions answered in SQL: revenue KPIs, RFM customer segments, cohort retention and team performance, each with recommendations |
| Sentimental-Analysis | Classifies the sentiment of text reviews to show how customers feel |
I'm open to roles and collaborations in business analysis, data analysis and digital transformation. Reach out on LinkedIn or by email.