Projects
Things I've built.
A few I'm proud of up top, then every public repo on my GitHub, newest first.
01 · FEATURED
Pothole classifier
A 16 MB model that spots potholes in road photos, fully offline on a laptop CPU, for a public pothole-reporting app where photos must never leave our machines.
| Model | Accuracy | Potholes caught | Speed |
|---|---|---|---|
| Mine (MobileNetV3, fine-tuned) | 89.5% | 84.5% | 3.4 ms, CPU |
| Ready-made HF pothole model | 65.4% | 35.4% | 13 ms, GPU |
| Clef-flash 9B, zero-shot | 61.1% | 18.9% | 389 ms, GPU |
| Google SigLIP, zero-shot | 58.2% | 13.1% | 31 ms, GPU |
- Merged 3 open datasets (South Africa dashcams, Kerala roads), removed 1,366 near-duplicates so the test set is honestly unseen.
- Fine-tuned MobileNetV3 in ~2 minutes; one harness scores any model on the same 607 held-out photos.
- Beat a 9B-parameter general model while running ~100x faster.
02 · FEATURED
Pia, the AI analyst in Mynt
At Pyvot I own Pia end to end: restaurant owners ask about their Zomato and Swiggy payouts in English, Hindi, Hinglish, Bengali or Marathi and get real numbers with an auto-picked chart.
Built at work, so the code is private.
- Every number in an answer is checked against the query results: unverified answers fell from 46% to 4%.
- A two-stage prompt-injection screen caught 85 of 85 red-team attacks with zero false flags on 1,082 normal questions.
- Moving pipeline steps to Claude Haiku cut cost per step by ~95% with the same decisions.
03 · FEATURED
ResearchAgent
A production-style AI agent built layer by layer in public: takes a research question, searches the web, reads sources, builds a knowledge graph and writes a structured report, remembering what it learnt across sessions.
- Built to understand every layer of the agent stack, not just call it.
- Evals and guardrails from day one.
04 · FEATURED
AskData
A conversational BI prototype: ask a business question in plain English, it writes the SQL, runs it, and answers with a summary, a chart and a table. Asks a clarifying question when the question is ambiguous.
All 35 repos
pothole-classifierOffline pothole / no-pothole image classifier: open data, fine-tuned MobileNetV3, 89.5% test accuracy, ~3 ms per image on CPU, with evals vs Clef-flash, SigLIP and an off-the-shelf model
family-treeAn interactive family tree you can share with a link: generations, how any two people are connected, search.
palette-palsWebsite for Palette Pals, an art duo doing murals, relief work, live wedding painting and workshops. Plain HTML/CSS/JS.
facility_usageLeakage-safe next-booking prediction for a residential community's gym, pool, courts and halls (take-home assignment).
sql-query-generatorAskData: ask business questions in plain English, get SQL, a summary, a chart and a table.
research-agentProduction AI agent with RAG, Graph RAG, MCP, evals and guardrails — built layer by layer
optimizer-appOptimisation app experiment.
mlops-testMLOps CI test repo.
mlops3MLOps practice, round 3.
mlop2MLOps pipeline practice, round 2.
mlopsMLOps pipeline practice.
SnackarooFlutter app that suggests healthy snacks for your craving: sweet, crunchy, salty or 'not sure'.
zariyaNext.js app experiment.
recodemindFlutter app experiment.
healthsageFlutter health app experiment.
story-writer-aiAn AI that writes stories.
deskfitChrome extension for desk workers: Pomodoro timer, stretch and water reminders.
ramayan-quizA React quiz on the Ramayan.
PrivaChat-Law-DemoPrivate, local chat over legal documents (demo).
sentiment-analysis-mlopsSentiment analysis model with an MLOps pipeline.
ai-code-doc-generatorPaste code, get documentation, generated by an LLM. Live on Vercel.
prettiest-wife-everA photo collage website made for my wife. 'There she is!'
juststopExpo / React Native app experiment.
llama2-ownpdfasking question from you files using llama2
Satellite_Image_ClassificationClassification of images captured by Satellite
House-Price-PredictionRegression models for house prices.
ValueInvestingValue-investing screens in a notebook.
BankLoanDefaultPredicting bank loan default.
NLP-Deployment-Heroku
Heroku-Demo
LDALDA on 20-newsgroups dataset
Data-Science-HacksData Science Hacks consists of tips, tricks to help you become a better data scientist. Data science hacks are for all - beginner to advanced. Data science hacks consist of python, jupyter notebook, pandas hacks and so on.
DataScienceProjectsEarly data science projects and notebooks.
Data-Science--Cheat-SheetCheat Sheets
Deep-Learning-CourseraDeep Learning Specialization by Andrew Ng, deeplearning.ai.