Pune, India | Data Science and AI Engineer | B.E. AI & Data Science '27 (CGPA 9.28) | Ex-Founder, MahaGuru AI (COEP I2I)
Building AI systems that reason, verify, and know when to stop.
An AI repair assistant for Indian households, in English, Hindi and Hinglish. Describe what is broken by typing, talking or taking a photo. Fixy asks a few targeted questions, names the likely fault with a confidence score, walks through a safe and sourced fix, and stops the session when the job needs a professional.
| Diagnosis engine | A Bayesian engine over a curated knowledge base (ceiling fans, extension boards, headphones, battery toys) picks the most informative next question. 90.7% top-1 and 96.0% top-3 on the 94-case FixBench-IN benchmark, versus 81.3% top-1 for plain Gemini. It keeps working when the LLM quota runs out. |
| Verified answers (RAG) | Hybrid retrieval on Supabase (pgvector + PGroonga full-text, Reciprocal Rank Fusion) over curated guides and Wikipedia, 100% recall@5. Gemini drafts with citations and Gemma checks every claim: 92.5% correct answers and 0% unsupported claims, versus 87.5% and 34% for plain Gemini. Says "not sure" when the sources do not cover a question. |
| Safety gate | Rules in three languages plus a TF-IDF + logistic regression classifier trained on 3.8K examples and exported to a 0.68 MB pure-Python runtime (about 0.2 ms per message). Hazard recall rose from 67.2% to 89.1%; safety-stop recall is 100% versus 84.6% for plain Gemini, with 0% false stops. |
| Product | Fixy, an animated robot guide; photo identification with location data removed on the device; voice input and spoken replies; step-by-step guides with tools, time, a rupee estimate, cited sources and part search links. |
| Engineering | FastAPI and Pydantic with an LLM provider chain and model fallbacks, per-IP rate limits, Docker on Render, a Next.js PWA on Cloudflare Workers, and CI with ruff, 118 pytest tests and a benchmark gate that fails the build if any safety case does not stop. |
flowchart LR
U["Text, voice or photo"] --> W["Next.js PWA<br/>Cloudflare Workers"]
W --> A["FastAPI on Render"]
A --> S{"Safety gate<br/>rules + classifier"}
S -- "hazard" --> X["Stop screen<br/>112 and electrician"]
S -- "safe" --> D["Bayesian diagnosis<br/>curated knowledge base"]
D --> G["Repair guide<br/>tools, cost, parts, sources"]
A --> R["Hybrid retrieval<br/>pgvector + PGroonga, RRF"]
R --> L["Gemini draft<br/>with citations"]
L --> V["Gemma claim check"]
V --> W
Benchmark report · QA report · Safety model card · Design
The API runs on Render's free tier, so the first answer can take about a minute while it wakes up. Benchmarks are small, author-written sets; the reports list their limits.
An open-source AI mentor and personalised classroom for college students. Rebuilt from my COEP I2I startup; live, tested and deployed.
| StudentGPT | A reflective mentor that asks before it advises. English and Hinglish safety screening on every message, a structured understanding record, and a clarity summary at the end. |
| Classroom | Turns a goal into a diagnostic, a personalised roadmap, generated lessons with an AI teacher, quizzes, graded projects and adaptive mastery tracking. |
| Engineering | FastAPI and React (TypeScript) monorepo, PostgreSQL, streamed replies over SSE, a provider-agnostic LLM layer (Gemini by default), one Docker image. |
| Quality | pytest on SQLite and PostgreSQL, Playwright end-to-end tests on desktop and mobile, CI on every push, and a 24-scenario evaluation suite including crisis cases. |
Research behind it · Architecture · Evaluation method · Screenshots
Projects from my resume first, with live demos wherever one is available.
| Project | What it does | Links |
|---|---|---|
| Customer Churn Prediction | Kaggle Playground S6E3. XGBoost and LightGBM on ~594K telecom records (21 features) with 5-fold stratified CV; XGBoost reached 0.916 mean ROC-AUC. EDA surfaced the strongest churn drivers: electronic-check payment, month-to-month contracts, fiber-optic internet and first-year customers. | Live demo · Source |
| CreditWise: Loan Approval | scikit-learn pipeline on 950 applicants (18 features). A class-balanced Random Forest reached 0.91 F1 and 0.98 ROC-AUC, versus 0.81 F1 for a logistic-regression baseline. Deployed with live predictions, confidence scores and feature importance. | Live demo · Source |
| IAMARS: Drone Detection and Tracking | Team project. YOLOv8n fine-tuned on VisioDECT with a leakage-aware split: 0.965 mAP@0.5 on a 1,800-image held-out set. ByteTrack tracking, 88.5 FPS on an RTX 2050, and intercept-point prediction in 3D simulation. | Live demo · Source |
| PhonePe Pulse Insights | Internship project. ETL of PhonePe Pulse JSON data (2018 to 2024) into a 9-table MySQL database, 10 SQL business questions, and a multi-page Streamlit and Plotly dashboard. | Live charts · Source |
| Mental Health in Tech: EDA | 1,251 survey responses analysed for what drives treatment-seeking: family history, work interference and employer benefits. Published as an interactive Plotly dashboard. | Live dashboard · Source |
| Tesla Stock Prediction | Deep learning pipeline: SimpleRNN, LSTM and a tuned LSTM with 1, 5 and 10-day forecasts. | Source |
| Real Estate Advisor | XGBoost model predicting property profitability and future value. | Source |
| Yes Bank Stock Price Prediction | Machine learning on historical OHLC data to predict Yes Bank closing prices. | Source |
| iTunes Music Store Analysis | SQL analysis of the Apple iTunes database for insights on customer behaviour. | Source |
More in my portfolio and on GitHub.
- Building Fix with Me — A bilingual AI repair assistant: Bayesian diagnosis, verified RAG answers and a trained safety gate
- Maintaining MahaGuru AI — A reflective AI mentor and an adaptive AI classroom for students, open source and live
- Evaluating LLM systems — Benchmarks against plain-LLM baselines, per-claim verification, safety recall and calibration
- Shipping ML pipelines end-to-end — From EDA and model selection to deployment and monitoring
- Researching human-centered AI — Presented at INMEC-2026, Keystone School of Engineering, Pune (reflective mentorship with LLMs)
Founder — MahaGuru AI (Jan 2024 – Nov 2025; rebuilt as open source in 2026) Founded an AI student-guidance platform, selected into COEP's I2I Startup Program for mentorship and incubation. Designed a multi-agent LLM architecture (FastAPI, React), curated a custom student-mentor dialogue dataset, and collected feedback from about 150 student users on early prototypes; the underlying research was presented at INMEC-2026. After the company shut down, rebuilt the product as an open-source platform that is now live.
Data Science Intern — Innovexis (Feb 2026 – May 2026) Built an ETL pipeline loading PhonePe Pulse JSON data (2018–2024) into a 9-table MySQL database, and a multi-page Streamlit + Plotly dashboard answering 10 business questions via SQL.
52 Weeks Contribution Activity
Statistics
Badges Earned
52 Weeks Submission Heatmap | Last Updated — 01/05/2026
Presented "AI-Powered Reflective Mentorship" at INMEC-2026 (1st International Conference, Keystone School of Engineering, Pune, Mar 2026)
B.E. — Artificial Intelligence & Data Science, APCOER, Pune (2023 – 2027, expected May 2027) — CGPA 9.28
HSC Science, Yashodhara Junior College, Solapur (2021 – 2022) — 78%
- Presented at INMEC-2026, Keystone School of Engineering, Pune — AI-Powered Reflective Mentorship
- Selected into COEP's I2I Startup Program — institutional backing for MahaGuru AI
- Certified AI/ML Engineer - Apna College Prime 2.0 Batch
Build systems that are useful, measurable, and honest. Ship working prototypes. Iterate from real feedback. Let the model do the heavy lifting — but understand every layer beneath it.
Random Facts
- Curated my first LLM dataset by hand: student-mentor dialogues for MahaGuru AI
- Founded an AI startup before finishing undergrad
- Competed on LeetCode to sharpen DSA fundamentals
- Believe in building small, shipping fast, and improving in production
- Based in Pune — powered by chai and curiosity