Made for stakeholders. Built for scale.
Pipelines, predictions, and agentic AI systems. I turn multi-dimensional data into decisions — and build the systems that keep doing it without me.
Not just models.
End-to-end systems that keep predicting without me.
01 / About
Dedicated Applied Data Sciences candidate combining rigorous academic training with practical experience in machine learning, data engineering, and AI system architecture. I design end-to-end predictive pipelines, automate enterprise data workflows, and run robust statistical analysis, including anomaly detection and A/B testing.
Proficient in Python and SQL/T-SQL, with a proven ability to translate multi-dimensional datasets into strategic business intelligence, optimizing operational efficiency and mitigating risk across technology, consulting, and manufacturing.
Weekly hours reclaimed by ETL automation at GRID
From a month of manual CSV joins to a scheduled pipeline.
Enterprise APIs wired into the Flying Sapphire agent
Google Workspace, Slack, and logistics tooling — all on-device.
Score on the A/B testing machine-learning assignment
Wine quality, hypothesis-tested and peer-graded.
End-to-end ML pipelines shipped
From raw source to decision and dashboard.
Three practice areas — 02 / Offerings
I connect, clean, and automate.
Teams drowning in weekly manual data prep.
I build predictions that hold up.
Risk, forecasting, and decision automation.
I make data legible.
Leaders who want decisions, not spreadsheets.
03 / Experience
Data Scientist
03/2024 – 12/2024
Amritsar, India
Executed comprehensive data cleaning, wrangling, and re-shaping on raw client datasets, establishing strict data-quality gates before any statistical work touched the data.
Cleaner inputs → trustworthy outputs
Architected and optimized SQL queries to extract, join, and relationalize multi-dimensional customer and sales data from enterprise local databases — replacing slow, fragile ad-hoc scripts.
Stable extracts at a fraction of the time
Engineered an automated ETL workflow in Python (Pandas) to consolidate disparate monthly CSV files into a single trusted source, eliminating manual processing.
≈ 5 hours reclaimed per week
Designed and deployed dynamic tracking dashboards to monitor KPIs, translating raw metrics into actionable insight for stakeholders across the business.
Built-in BI for decision makers
Partnered with team members to communicate complex analytical findings, driving strategic decisions through clear, evidence-first storytelling.
Findings → decisions, not reports
04 / Selected work
Four products, each taken the whole way — from a messy first source to a decision someone could act on.
Agentic AI Web Browser — Co-Founder & AI Workflow Developer
A browser that thinks. Local LLMs route tasks, read active tabs, and execute multi-step work across 40+ enterprise APIs — with no cloud computation.
The problem
Typical AI assistants ship your context to the cloud and stop at chat. Sapphire had to act on what you're actually looking at — live tabs, live data — and do the work end-to-end, privately, on the machine itself.
Approach
Integrated local LLMs and engineered the agent's reasoning loop so prompts are parsed into intent and routed to the right tool with no cloud round-trip.
Built parsing pipelines that distill raw, unstructured web text into clean, structured Markdown study guides and structured JSON.
Wired the agent to 40+ external enterprise APIs — Google Workspace, Slack, logistics tooling — for autonomous logistics and data-entry tasks.
Tuned system prompts and context windows so the agent reads live tabs, debugs Python, and produces insights at near-zero latency.
Outcomes
All routing and decisioning runs on-device; privacy is the default, not a setting.
A single natural-language trigger can span calendars, docs, and messaging.
— The browser that brings its own hands.
Forecasting & Anomaly Detection Pipeline
An end-to-end ML pipeline that forecasts match outcomes and flags irregular betting patterns — a simulated automated fraud and risk detection stack.
The problem
Bookmaking data is messy, interdependent, and adversarial. The task was to build the whole risk stack from scratch: forecast outcomes, then catch the bet patterns that do not behave like the model expects.
Approach
Built a Python + Scikit-learn pipeline to forecast match outcomes from historical transactional data, engineered for drift rather than best-case accuracy.
Used Pandas and NumPy to clean raw betting datasets, impute missing values, and engineer rolling statistical features.
Implemented classification models to surface high-risk betting anomalies and flag them for automated review.
Structured and queried T-SQL tables to store and process multi-dimensional match and odds metrics efficiently.
Outcomes
Ingestion → features → forecast → risk flag, tracked through MLflow.
Designed so a real-time odds feed can be dropped into the pipeline unchanged.
— Predict the game. Police the bet.
Predictive Modeling & A/B Testing
Rigorous A/B testing paired with a predictive pipeline classifying wine quality metrics. Scored a perfect 100% as a university assignment.
The problem
A deceptively simple dataset hides a demanding question: which chemical properties actually move quality, and which differences are just noise? The brief demanded statistical honesty, not a leaderboard win.
Approach
Extracted, joined, and preprocessed structured datasets from both CSV and JSON sources, keeping provenance intact.
Designed and executed A/B tests to statistically validate hypotheses on chemical properties and their effect on quality scores.
Documented the complete ML lifecycle — cleaning to evaluation — in Jupyter Notebooks for reproducibility.
Outcomes
100% on the assignment — the validation that A/B rigor transfers to production behavior.
Every figure in the report can be regenerated from the notebooks.
— Statistics can taste too.
Classification & Hyperparameter Tuning
A classification pipeline that automates credit approval decisions from historical financial data, tuned to cut false positives.
The problem
Credit data is brutally unbalanced and full of missing fields. A naive model over-approves and bleeds risk; an over-tuned one quietly rejects good customers. The brief: automate the decision without inheriting either failure.
Approach
Cleaned the highly unbalanced financial dataset by imputing missing values, encoding categoricals, and scaling features for stability.
Used GridSearchCV to optimize Logistic Regression parameters, prioritizing precision and reducing false positives.
Outcomes
Hyperparameters chosen against false positives, not raw accuracy.
Pipeline structure that ports directly to other binary approval tasks.
— Fewer bad yeses. No good noes.
Interactive — 05 / Playground
Every threshold is a trade. Move it left and you catch every anomaly plus half the honest traffic. Move it right and the quiet fraud slips through. This is the exact judgment a risk model encodes.
Anomaly detection — live demo
Research note — 06 / Method
Paper — internal note
A personal methodology for data products
Abstract
Guided by coursework in applied data sciences and hardened by production work at GRID and on agentic systems, I treat every engagement as a pipeline, not a script. Problem framing precedes acquisition; evaluation precedes deployment; and stakeholder legibility is a first-class output, never an afterthought. This note records the sequence I run — and the failure modes it is designed to prevent.
cite this method
@misc{khanna2026pipeline,
author = {Khanna, Arjun},
title = {From Raw to Rendered: A Personal Methodology for Data Products},
year = {2026},
school = {MCAST — Applied Data Sciences},
note = {Compiled from GRID production work and Level 6 coursework}
}01
Decide what decision this data is meant to support before touching a value. A model without a downstream decision is a technical exercise.
02
Find every source — CSV, JSON, SQL, API. Join them deliberately, gate data quality, and make provenance reproducible. This is where pipelines are won or lost.
03
Explore to build hypotheses, then model to test them. Choose the simplest model that clears the bar; reserve complexity for where it earns its keep.
04
Track everything in MLflow. Prefer precision and business-relevant metrics over leaderboard accuracy. Iterate on the metric the stakeholder feels, not the one the notebook loves.
05
Rendered output — dashboard, brief, or agent — that a stakeholder can act on. Then automate the pipeline so the next refresh does not need a human babysitter.
07 / Capabilities
01
02
03
04
05
06
07
08
Education & training
Applied statistics, machine learning, data engineering, and decision science — studied in the lab and pressure-tested on real pipelines.
6
MQF/EQF level 6 (honours)
100%
A/B testing ML assignment
4+
Models taken to production
Language skills
| Language | Listen | Read | Speak | Interact | Write |
|---|---|---|---|---|---|
| HindiMother tongue | — | — | — | — | — |
| EnglishProficient user | C2 | C2 | C2 | C2 | C2 |
Levels: A1/A2 basic user · B1/B2 independent user · C1/C2 proficient user.
08 / Contact
A pipeline, a prediction, a dashboard — or a browser that does the work for you. If the data exists, there is a way to make it decide. Let's find it.