I take the vague version and ship the real one.
Jay Hemnani. Forward Deployed Engineer and Software, Data and ML Engineer from Gujarat. Data pipelines, ML, agents, and occasionally a Rubik's cube. Currently open to work and, apparently, to building toys for my own homepage.
Today's pick
changes in --h --m --s
project of the day
Data Clean Env
THE BRIEF
An OpenEnv reinforcement-learning environment for tabular data cleaning: an agent fixes messy datasets through typed actions, scored per issue resolved.
WHAT CHANGED
Turns a fuzzy chore into a checkable RL task: each fixed issue is worth a fraction of the score, so the reward tracks real data-quality improvement rather than activity.
- Python
- FastAPI
- OpenEnv
- Docker
useless fact #7
Favourite bug: the one that only appeared on Tuesdays.
- projects, each with a write-up
- 28
- pull requests merged upstream
- 202
- IEEE papers (2021)
- 2
- fraud-model precision on transaction data
- 94%
- Rubik's PB, officially
- 16.7s
Things I've shipped
Six of twenty-eight. The other twenty-two are shy.
- 01FIFA Soccer DS“Match footage and a question: can we track every player and turn it into tactics?”Detection and tracking at 22 fps on my RTX 3070 laptop. Reproducible from extracted frames with one command.YOLOv8ByteTrackDVCMLflow
- 02CAG Deep Research“Research agents that confidently agree with themselves.”Every claim exits tagged VERIFIED, FALSIFIED or UNCLEAR, with the evidence.LangGraphLangChainOllamaGroq
- 03Stock Data Platform“Tick data nobody could query and a dashboard everyone wanted.”Multi-symbol OHLC from ingest to a Chart.js dashboard; a warehouse a BI tool can join.KafkaAirflowTimescaleDBChart.js
- 04WebMCP, on this site“Could a portfolio be read by a machine as easily as by a person?”An agent in Chrome, with the WebMCP flag on, reads this site without scraping it.TypeScriptNext.js 16WebMCP
- 05Biotech Accelerator“A week of tab-hopping across UniProt, PDB and ChEMBL for every hypothesis.”Question in, ranked candidates and citations out, in minutes.LangGraphProDyhttpxDocker
- 06Nobel Data Intelligence“Protein stability prediction stuck on sequence alone.”Still predicts when no structure is available. Whether the vibrational branch helps is not benchmarked yet.PyTorch GeometricProDyRDKitTransformers
House rules
Each one cost me a project to learn. You get them free.
PERSONAL BEST · WCA OFFICIAL
16.7s
Click the cube to scramble it. I promise I'm faster than this.
- #1
Boring parts first.
Ingestion, schema, tests. The clever layer earns its place once the dull one holds.
learned on: Stock Data Platform
- #2
Re-runnable, then better.
If a result can't be reproduced with one command, it isn't a result.
learned on: FIFA Soccer DS
- #3
Argue with the model.
Agreement is not evidence. Give it an adversary and a judge.
learned on: CAG Deep Research
- #4
Publish the gap.
82.68% in the paper, 73.59% in the notebook. Both numbers live on the résumé.
learned on: IEEE 2021
Where I've been
Summer 2025
Data Analyst @ Elite Hotel Group
ETL in SQL + Python that cut manual prep 40%; occupancy and revenue dashboards; time-series demand forecasts for pricing.
2022-2024
Technical Consultant @ Independent
Analytics and pipeline work for small businesses; A/B frameworks; reporting automation.
Jan-May 2022
AI/ML Intern @ Amnex
Credit-fraud ensemble with SMOTE, 94% precision; anomaly dashboards.
Jun-Aug 2021
iOS App Development Intern @ Cygnus SoftTech
CodeLock: AES-encrypted iOS privacy app on Core Data.
May-Nov 2019
SWE Intern @ Cactus Creatives Pvt. Ltd.
First-responder comms platform on Azure microservices; CI/CD cut deploys 60%.
Leave a sticky
counting notes…