Yeshwant Singh

Music Subscription Web App (Serverless on AWS)

I got tired of just reading about cloud architecture, so I built something. This is a full-stack music subscription app where users can register, log in, search a catalogue by title, artist, or year, and curate their own library (all running on a serverless AWS setup that fits comfortably within the free tier)

The backend lives in Python on AWS Lambda, which means functions only spin up when they're actually needed. No idle servers, no babysitting infrastructure. API Gateway sits in front and gives the frontend clean REST endpoints to talk to. DynamoDB handles the structured data (user accounts and music records), and artist artwork is served straight from S3. The frontend is plain HTML, CSS, and JavaScript, deployed on GitHub Pages (so the whole thing, UI and cloud backend, is live and shareable)

Tech stack: AWS Lambda · API Gateway · DynamoDB · S3 · Python · HTML/CSS/JavaScript · GitHub Pages

The part I found most valuable wasn't any single service; it was tracing a request all the way from a browser click, through API Gateway, into a Lambda function, and back out from DynamoDB or S3. That end-to-end picture is hard to get from documentation alone. I also cared about the details that make a project feel finished: a responsive UI, sensible data handling, and a setup that could actually scale without me having to think about it.

Website: https://singhyeshwant.github.io/music.subscriptions/

This project built an interactive dashboard (using R-programming) that shows different layers of social and economic data across Australia. The main goal was to create a simple way to see information about health, education, jobs, and justice, especially looking at differences affecting Indigenous communities.

The dashboard was made using R Shiny, a tool that helps create easy-to-use and flexible interfaces. It includes charts and tables that users can change to look at different states and years. There is also an option to download the data, making the dashboard useful for many people like policymakers and researchers.

One important part of the project is the attention to small details, such as showing trends in healthy birthweights and examining youth involvement in education and jobs. Each chart comes with a short explanation that helps explain what the data means, showing what has improved and what still needs work. These explanations help tell the story behind the numbers and highlight where special efforts might be needed.

This dashboard is more than just a way to see data. It aims to inform and teach people, helping to start conversations, encourage change, and support steps toward a fairer society

Website: https://yeshwant.shinyapps.io/3994442YeshwantSingh/

Created an NLP-based predictor to detect harmful undertones in text using ensemble modeling. Trained and evaluated the model against a large set of test sentences to achieve high classification accuracy. Packaged the outcome for stakeholder presentation within the RMIT context.

https://rmit-arc.instructuremedia.com/embed/269fc476-53de-4a74-851b-6377b7dd465b

I manage several multilingual AI projects related to annotation, transcription, validation, review, audio/video evaluation, etc. The majority of my role involves interactions between clients, internal colleagues and resources, where I need to make sure that project details, current task updates/blockers, and future follow-ups are properly tracked.

Here, I

• Manage 8+ AI data projects in different languages, task kinds and delivery dates at the same time.

• Build Power BI dashboards pulling live data from Sheets and Asana to visualise throughput, quality trends and resource bottlenecks across all active projects, reducing weekly report preparation by 6+ hours and surfacing at-risk deliveries 2-3 days earlier.

• Coordinate various projects involving 25+ languages, which encompasses the following: availability checks, onboarding, eligibility monitoring, task updates and resource follow-ups.

• Use Google Apps Scripts, Google Sheets and Docs to create automation workflows that allow me to track the pending volume of work through project readiness, assignment of tasks, invitation to platforms and progress of the resources.

• Establish tracking and follow-up systems for overall easier monitoring of pending work. Technically, these efforts resulted in an estimated increase of approximately 50% - 55% in the rate of project completion.

• Help both clients and internal teams with confirmation of project scope/language needs/AHT/guidelines/expected volume/start date/blockers prior to project implementation.

• Prepare handover notes for teams on what has been completed and what is left to do.

In this role, I worked on AI training data duties such as annotation, validation, SFT, RLHF and quality review procedures. The work involved mostly validating task outputs against client instructions and carefully reviewing model replies. I've also done client projects for firms like Uber and NVIDIA and some post-training work on the Translated's Lara AI model.

• Managed and processed more than 1,000 AI training jobs in annotation, validation and quality review activities.

• Performed SFT and RLHF-related activities to contribute to the post-training of Lara, Translated's AI translation model, trained on NVIDIA AI platform and GPU infrastructure.

• Verified model outputs for accuracy, instruction match, language quality and overall task correctness.

• Raised ambiguous guidance and repetitive task-level issues early so they could be addressed before they could affect more work.

Developed an AI-powered patient triage tool that helped doctors to examine patient summaries way faster than usual and notice important cases earlier. The majority of the effort consisted of translating the patient's descriptions into a simple priority score so that doctors could understand which cases needed care most urgently.

Here I've

• Developed a patient triage system which sorted 1,000+ patient summaries and cases depending on urgency.

• Developed keyword detection and scoring logic to identify medical red flags from patient descriptions.

• Developed a priority score for patient text that might be used for faster review by clinicians.

• Reduce manual report scanning by an estimated 97% with the earlier identification of critical patients and key clinical markers.

• Helped design a technology that could assist doctors in spending less time perusing extensive reports and more time viewing high-priority cases.

Here I've trained a data preprocessing AI model by reviewing/rating and refining its ability to efficiently handle tasks involving CSV/Excel/etc files. Precisely, I've evaluated the AI model's background step by step approach, which includes dataset cleaning, file merging, attribute removal and much more (to a logical progression towards the final output). I've completed 12+ high quality AI training tasks per day (with tasks taking ~ 2 hours each) using Jupyter Notebook, Python/pandas/NumPy/openpyxl/xlrd/scikit-learn preprocessing tools and regular expressions for data analysis.

I've also worked on another project during this time, where I examined AI coding models for their responses to code explanation, debugging, and code generation. Verified 500+ solutions in Python, C++, JavaScript and C for accuracy, usefulness, clarity and job match. Checked corrections of wrong codes; made solutions and explanations; provided structured comments for correct, incomplete or unclear responses.

Here I:

• Worked with Production, Quality Assurance and Defence Compliance to work out what we actually needed to track on the manufacturing side, then built that into automated quality dashboards. Defect reporting time came down by 35% once these were running.

• Pulled data together from ERP systems, defence contract platforms and equipment testing logs using SQL. Every month that meant consolidating 10,000+ records across production, inventory and calibration into something the team could actually use.

• Built and kept up Power BI dashboards for what we called our Key Defence Metrics: equipment readiness, calibration compliance against ISO and defence standards, and on-time delivery on the critical GSE contracts. Stakeholders could check where things stood without waiting on someone to send a report.

• Dug into production delays and equipment failures at one point, and traced 23% of the hydraulic jack defects back to a single supplier batch. That led to corrective action from the supplier and saved around ₹4.2 lakh (A$7,500) a year in rework.

• Put together forecasting models for spare parts demand on the SU-30 MKI ground support equipment. Inventory turnover improved by 18%, and we still kept a 99.2% service level for contract fulfilment.

• Took defence compliance reporting, monthly and quarterly, from 16 hours down to 2 hours per cycle by automating it, and got rid of the data reconciliation errors that used to come with the manual process.

• Ran A/B tests on calibration equipment maintenance schedules and used what we learned to recommend predictive maintenance intervals. Equipment uptime went up by 12% and unplanned downtime costs dropped.

• Ran a segmentation analysis on our defence clients and found that a small group of high-priority contracts made up 65% of revenue. Acting on that helped lift client retention and pushed contract renewals up by 14%.