> hello_world.py
I'm Kashish, and I'm
a data analyst.
I make messy data tell the truth, and I study when models quietly don't. Data analyst and research assistant. Finance, healthcare, applied AI. MS Data Science, Stony Brook.
photo: real. the sunset was not A/B tested.
> whoami
The analyst who checks twice.
I like the part of data work most people skip: making sure the numbers are right before anyone bets a decision on them.
I've reconciled factory data from systems that couldn't agree on what to call a machine, automated reports a finance team lived in, run A/B tests at a startup, and measured how much search rankings wobble when you reword a question. AI is in my toolkit too. I just make it show its work.
My love language is a reconciled total.
SELECT * FROM kashish;
1 row returned · 0 errors ▍
> git log --career
The commit history.
Squash-merged for your convenience. No force-pushes were harmed.
Research Assistant
CUBIT Lab, Stony Brook University · Dr. Christine DeLorenzo, USA
Health Research Data · Data Quality
- Onboarding to the actigraphy quality-control workflow for health research data.
Research Assistant
Stony Brook University · Prof. Rahmati & Kevin Eykholt, USA
Information Retrieval · Robustness · Evaluation
- Asked whether search rankings stay put when you reword a question without changing what it means. (Spoiler: they wobble.)
- Built a reproducible pipeline with cached results to compare Google and DuckDuckGo on the same reworded queries.
- Defined True-PIR, a demotion-only stability score, plus bias diagnostics: Google ≈ 0.625 vs DuckDuckGo ≈ 0.586 (subset).
Founding AI Engineer
Versaunt (fka. Aayats), Inc., USA
Product Analytics · Experimentation · Quality
- Owned analytics, quality and monitoring at a small AI startup.
- Ran A/B tests and cohort analysis across 50K+ daily events.
- Built the dashboards that tracked a platform at 10K+ users and 500K+ queries.
- Shipped a few small features. The engineering team built the core backend, credit where it's due.
Data Analyst
Saroj Group of Companies, India
Finance Reporting · Forecasting · Automation
- Ran forecasting and cost scenarios that supported a 12% margin gain.
- Built KPI dashboards in SQL, Excel and Tableau that cut reporting time by 35%.
- Automated four Python and SQL reporting pipelines, giving the team back 25+ hours a month.
- Used statistical modeling on 100K+ records to help cut defect rates by 40%.
Data Analyst
Kamala Plastics, India
ETL · Reconciliation · Operations Analytics
- Got 200K+ records from systems that couldn't agree on what to call a machine into one standard.
- Reconciled counts and totals until the reports finally matched.
- Ran root-cause analysis that gave engineering the evidence to cut scrap by 22% and returns by 15%.
Research Assistant
Dept. of AI and ML, D.Y. Patil College of Engineering and Technology, India
Applied AI Research · Mentorship
- Built a CNN-based malaria detector from blood smear images (92% accuracy).
- Built a real-time YOLOv4 object detector and an LSB image steganography method.
- Led the research track for a Smart India Hackathon national finalist project (Ministry of Power problem statement).
- Mentored peers as Head of R&D for the AIML Students Association.
> ls ./projects
Things I shipped.
Click one. They all come with receipts.
// Research
- problem
- Do search rankings stay stable when you reword a question without changing what it means?
- what_i_did
- Built a reproducible pipeline with cached results comparing Google and DuckDuckGo on reworded queries. Defined True-PIR, a demotion-only stability score, plus bias diagnostics.
- result
- True-PIR ≈ 0.625 for Google vs ≈ 0.586 for DuckDuckGo (subset). Same question, different rankings.
- tools
- Python, statistical testing, Google CSE, DuckDuckGo
- what_i_learned
- A metric only counts if someone else can rerun it and get the same number.
Written up as a Medium article, not a paper.
// Finance and business analytics
// Health data
// Applied AI and ML
> config.defaults
How I work.
validate()
Check the data first
I check the data before anyone sees a chart. Trust issues, but make it professional.
reconcile()
Make the numbers add up
Rows in should equal rows out. If they don't, nobody goes home.
document()
Write it down
I write it down so future me, and future you, don't have to guess.
explain()
Say it plainly
Decision first, jargon never. Okay, rarely.
flag_risk()
Flag risk early
If something is on fire, you hear it from me, not from the dashboard.
> import toolkit
What I work with.
Plus an unreasonable amount of oat milk.
Analytics and BI
Data Engineering
AI and ML
AI-Assisted Development
> education
Stony Brook University
Master of Science in Data Science
Aug 2024 to May 2026
Shivaji University
B.Tech in Computer Science (AI and ML)
Aug 2020 to May 2024
Best Outgoing Student (sole recipient)
> cat ./writing
I write things down so I don't forget.
Project write-up
The Threshold Is a Business Decision, Not a Statistical One
Why choosing the right cut-off in fraud detection matters more than accuracy alone.
Read on Medium →Project write-up
7 Years. That's How Long Rare Disease Patients Wait for a Correct Diagnosis.
On the rare disease diagnosis gap and how structured clinical data can help.
Read on Medium →Project write-up
Same Question, Different Words: Do Search Rankings Stay Stable?
Measuring how much rankings move when a question is reworded.
Read on Medium →Project write-up
Scaling Retrieval Systems: Why Smarter Memory Might Beat Bigger AI Models
Indexing choices, latency bottlenecks and what to fix first.
Read on Medium →Ideas
They Made It Hard to Leave. I Wanted to Know Why.
A data-driven look at the retention patterns that keep users around.
Read on Medium →Ideas
You're Not Prompting AI. You're Controlling a Probabilistic System.
A simple mental model for working with LLMs.
Read on Medium →> achievements.unlocked
Wins, and the receipts.
// leadership_and_recognition
- Best Outgoing StudentSole recipient · 2024
- Harvard HackRare FinalistClinical notes pipeline · team of 3 · 2026
- Smart India Hackathon FinalistNational, Ministry of Power · 2023
- Student Council SecretaryThree terms, about 4,000 students
- Best Emerging RotaractorAcross 75+ clubs · District 3170 Secretary of the Year
- District Chess Champion2023 and 2024
// certifications
- Claude 101 ↗2026
Anthropic
- Essentials of Prompt Engineering2026
Amazon Web Services
- AWS Academy Data Analytics ↗2023
Amazon Web Services
- Supervised ML: Regression and Classification ↗2023
Stanford University · Coursera
- Quantitative Finance Workshop2023
E-Cell, IIT Bombay
- Python for Data Science, AI and Development ↗2022
IBM · Coursera
> ping kashish
Got messy data? Send it my way.
Hiring for data or analytics, working on robust ML, or just want to talk numbers? I reply faster than your nightly ETL job. Yes, even if the attachment is called final_final_v3.xlsx.