Data analysis
SQL with CTEs and window functions, Power BI and DAX. Three dashboard suites for 8 fund managers.
Open to remote roles · 6 years in data
I work as a data scientistML engineerdata analystbackend developeragentic AI engineercivil engineer
I build Python and SQL pipelines, predictive models and Power BI dashboards. At Deccan Holdings they handled 4.5M records a day and cut a 6-hour report to 45 minutes.
Survey mode · find the 12 project sites
0 / 12 surveyed
The story · 2016 → now
Jawaharlal Nehru Engineering College (MGM). Structural design, estimation and surveying: my first models were made of concrete.
Intern at CBRE, checking sites against bar bending schedules and drawings. At Teemage Builders I made precast tracking about 20% faster.
A PG diploma: scheduling in Primavera P6, bills of quantities, QA and QC. I still read a BOQ as well as a query plan.
BI and MIS at Deccan Infrastructure recovered INR 38 lakh from billing. Then healthcare, finance and AI: pipelines, models and dashboards.
What I do · four disciplines
SQL with CTEs and window functions, Power BI and DAX. Three dashboard suites for 8 fund managers.
Classification, regression, time series and NLP, tuned to the cost of a wrong call. PyTorch for quantization, video search and deepfake detection.
Scheduled Python jobs that ingest, validate and publish, with tests and CI on every repository. Flask APIs, SQLite and Parquet stores.
HTML, CSS and React for pages like this one, and Streamlit apps on top of each model so people can actually use it.
Lab · try it yourself
Each dot is one of the 374 customers who churned in the test set. Lit dots are the ones the model flagged in time. Lowering the cutoff to 0.38 trades a little precision for 37 more saved customers.
SELECT contract, AVG(churned) AS churn_rate FROM customers GROUP BY contract ORDER BY churn_rate DESC;
3 rows · ran on 7,043 customers
Agentic AI · LLMs that use tools
I build LLM agents that break a goal into steps, call tools such as SQL, search and code, and check the result before they answer. Then I measure them: my eval harness gates every release, and at Outlier and micro1 I rate agent replies against rubrics.
Typed tools for SQL, search, files and APIs, with validation and retries.
Plan, act, observe and repeat within a step budget, with memory between steps.
Grounding answers in documents and video: the video-search project finds the moment that matches a sentence.
A 60-task agent benchmark with bootstrap intervals and McNemar tests. Candidate v2: pass rate 80.0% → 95.0%, p = 0.0039.
Rubric-based rating of LLM and agent replies in Hindi and English, with a written reason for each.
Real tools over this site's data. The planner here is a small rule-based router so you can watch every step; in production that role is played by an LLM.
Selected work · 12 repositories
Every number below comes from the repository's own tests. Scroll sideways through them.
All repositories on GitHubCUPED cuts the sample an experiment needs by 40%, and a sequential SPRT lets people check results every day without inflating false positives. Tested on 300 simulated experiments.
Weibull adstock, Hill saturation and a non-negative Bayesian ridge, then SLSQP moves budget between channels: +1.09% weekly revenue for the same spend, over 10 simulated 156-week runs.
BG/NBD and Gamma-Gamma models forecast 12-month value for 1,500 accounts, with Markov multi-touch attribution and Kaplan-Meier retention curves.
Multi-echelon safety stock for 1 warehouse, 4 hubs, 16 stores and 20 SKUs, plus a 52-week simulation of how order swings grow upstream.
Sets surge prices across 8 city zones and 96 fifteen-minute slots, and holds 1.0× when supply is balanced. Compared with flat pricing over 10 simulated days.
A churn model on 7,043 telecom customers. Moving the cutoff from 0.50 to 0.38 catches 307 of 374 churners in the test set.
Forecasts for 120 tickers in a SQLite and Parquet store, with checks for splits, bad ticks and gaps. The morning report runs in under 45 seconds.
A star schema with 40,000 sales facts, 12 KPIs in SQL and DAX, and drill-down by time, region, product, segment and channel.
A 60-task benchmark for LLM agents with bootstrap intervals, permutation and McNemar tests, and a CI gate that blocks a weaker release.
INT8 per-channel and INT4 group-wise quantization for a 12.86M-parameter model, with an FP16 softmax overflow test.
Finds the moment in a video that matches a sentence: shot cuts, a 512-dimension dual encoder and a cross-attention re-ranker.
Spots AI-made faces by pairing an SE-ResNet image stream with the 2D FFT power spectrum, which generators get wrong.
Experience · since 2017
beforeafter
Hindi and English audio, TTS and agent replies rated against rubrics.
Video annotation, rated senior.
Pipelines for 4.5M records a day; a 6-hour report down to 45 minutes.
12 US client networks, HIPAA audits.
450,000 frames checked.
INR 38 lakh recovered from billing.
Precast tracking, about 20% faster.
Site checks against BBS and drawings.
Toolbox · 14 tools, 4 languages
Drag the nodes around. Pick a discipline to light up its tools.
MGM's Jawaharlal Nehru Engineering College
NICMAR
Coding Ninjas certificate
Terminal · SQL over my career
Next · remote roles and contracts