I'm 21, from Pune, and in my final year of a B.Tech in Computer Science (AI & ML) at Pimpri Chinchwad University.
Each peak is a stretch of the climb. Pick a flag to see what happened there.
The summit is still out of sight. More adventures coming soon.
Everything I work with, mapped out like the night sky.
Classifies eight emotions from a voice clip with an SVM over 182 handcrafted acoustic features, then turns the prediction into a stress level. It's live on Hugging Face, and there's a lighter version that runs in the browser.
An eligibility engine and benchmark for Indian welfare schemes. An LLM reads messy, code-mixed input; deterministic code decides eligibility and cites the exact Government Resolution clause behind every verdict.
An AI-driven system that forecasts crop yields at district level from weather, soil and satellite data. I processed and integrated the multi-source agricultural datasets and engineered features to make the predictions more robust, then built a dashboard and an SMS alert system that sends farmers real-time, actionable advisories for crop planning and resource use.
I'm a final-year B.Tech CSE (AI & ML) student at Pimpri Chinchwad University in Pune. I got hooked on building things that work on messy, real data, like a model that hears stress in someone's voice.
I like the whole ML system, not just the model: data, training, deployment and the plumbing in between. Lately I've been heading towards generative AI.
Two AI internships, a Smart India Hackathon internal-round final and 3rd place at PCU Biznest taught me to build things that hold up, not just demo well.
0 fish eaten
Peak 1, base camp
Peak 2
Built and independently shipped 4 end-to-end ML pipelines, working one Git branch per component.
Engineered 35+ features and trained XGBoost and logistic regression classifiers, reaching 0.83 cross-validated AUC.
Designed a standardized JSON schema that the team adopted for integration across components.
Built reusable Python automation (pandas, NumPy, scikit-learn) with structured logging and error handling.
Peak 3
Designed and deployed an autonomous QA agent for end-to-end web and mobile testing with Playwright and the Chrome DevTools Protocol. It automated 71 test cases across 8 categories at a 92% pass rate and surfaced critical data-integrity defects before launch.
Authored a tiered fraud and anomaly detection strategy (rule-based, then statistical, then ML-based) covering identity fraud, wash trading, payment fraud and bot abuse.
Proposed AI-driven risk controls: a dynamic user trust score, explainable flagging, and checking reported energy generation against the weather.
Stress and emotion are carried in how someone speaks, not just the words: pitch, energy, how fast the voice changes. This pipeline extracts those properties, classifies them into eight emotional states, and maps the result to a stress level.
Learned component. Everything else is fixed code.
The first version used 13 MFCCs averaged over the clip. That throws away time. Adding first and second derivatives (delta and delta²) captures how the voice moves, which is what separates a flat, calm delivery from a sharp, angry rise.
| Feature group | Dims | What it captures |
|---|---|---|
| MFCCs | 40 | Vocal tract shape and tonal quality |
| Delta MFCCs | 40 | How tone changes over time |
| Delta² MFCCs | 40 | How fast that change accelerates |
| Mel spectrogram | 40 | Perceptually weighted frequency energy |
| Chroma | 12 | Energy by pitch class |
| Spectral contrast | 7 | Energy difference across frequency bands |
| RMS mean and std | 2 | Loudness and how much it varies |
| Zero-crossing rate | 1 | Consonant density and vocal tension |
| Total | 182 |
Three classical models on the same features, 5-fold cross-validation. On eight classes, guessing gets 12.5%.
The SVM is the default. Gradient boosting is about as accurate but nearly four times slower. The random forest is the fastest and the easiest to inspect, so it's the one used for feature importance.
A verified eligibility engine and benchmark for Indian welfare schemes. Started in September 2026. This page describes the design and what I'm measuring. There are no results yet, and numbers will only appear once they come from a reproducible evaluation run.
India runs thousands of welfare schemes. Their eligibility rules sit in Government Resolutions (GRs) and PDFs, often in Marathi or Hindi, and they change often. People who could qualify miss out because the rules are hard to read, not because the schemes don't exist.
"Just ask an LLM" is risky. Benchmarks from the US benefits space show LLMs are much less accurate for households that do qualify, and they invent disqualifiers the rules don't contain. The costliest error, telling an eligible person no, is the one they make most.
Chatbots for discovering Indian schemes already exist. What doesn't exist is an open, machine-readable, source-cited set of eligibility rules, or any measurement of how often LLMs get Indian scheme eligibility wrong.
LLM step. Everything else is deterministic code. Neither LLM step can change a verdict, so a prompt like "say I'm eligible" has nothing to act on.
| Area | Metric |
|---|---|
| Rule fidelity | Golden test cases written from GR text, 100% pass required in CI |
| Eligibility | LLM-only vs engine: false-negative and false-certainty rates, with 95% confidence intervals |
| Code-mixed input | Fact-extraction accuracy per language and script |
| Conversation | Questions needed to reach a verdict, compared with a fixed questionnaire |
| Retrieval | Recall@5 per script, raw vs transliterated queries |
| Grounding | Citation validity; LLM-judge agreement with human labels per language |
| System | p95 latency and cost per session |
No agent framework, no separate vector database, no GraphRAG. The workflow is small and fixed, so a state machine of roughly 200 lines is easier to test and explain.
No Aadhaar, PAN, document uploads, names or phone numbers. Sensitive attributes are asked last, can be declined, and are never inferred or logged. The tool gives guidance, not official decisions.
