Skip to content

Open to analytics roles in the GTA & Southwestern Ontario

Analytics that ends in a decision,not a dashboard.

Master of Management Analytics from Queen’s. I work at the point where optimization, machine learning and automation meet a real P&L — including my own: a home-goods brand whose inventory I plan with a linear program I wrote.

hassamasghar94@gmail.com

What I work with

  • Git
  • Agentic workflows
  • R
  • Claude & OpenAI APIs
  • Power BI
  • Regression & regularization
  • Linear & mixed-integer programming
  • pandas
  • Advanced Excel
  • Executive decks
  • Python
  • Looker Studio
  • A/B testing
  • scikit-learn
  • statsmodels
  • XGBoost
  • Hypothesis testing / ANOVA
  • Tableau
  • REST / GraphQL APIs
  • Data storytelling
  • PuLP / Solver
  • Simulation

Never all of it at once — the list keeps rotating.

Where the work comes from

  • Smith School of Business

    Queen’s University · MMA

  • Cowlar Design Studio

    Y Combinator–backed

  • Modisoft Inc.

    200+ retail accounts · POS data QA

  • Hopefield

    my own brand

What I build it with

Python
Power BI
Advanced Excel
pandas
scikit-learn
NumPy
SciPy
Plotly
Jupyter
R
Node.js
Git

Selected work

The work, grouped by the kind of problem it solves.

Every one started with a problem someone actually had — question, model, decision, result. Jump to whichever kind you care about.

Optimization

Linear and mixed-integer programs — where the answer has to survive a real constraint.

OptimizationHopefield — my own brand

Kitting 7 SKU families without stranding inventory

A linear program that turns loose towel stock into sellable retail sets — with almost nothing left over.

My home-textiles brand receives towels as loose pieces and sells them as bundled sets. Packing by intuition strands odd inventory that can never be sold. I replaced the guesswork with a linear program that maximises how much stock becomes revenue.

decision variables across the model
28decision variables across the model
line–colour families planned at once
7line–colour families planned at once
towel types balanced simultaneously
3towel types balanced simultaneously
Linear ProgrammingExcel SolverPythonInventory Planning
Read the case study
Set of 6 (combo)Set of 6 hand

Set compositions the LP packs to — bath, hand, washcloth

OptimizationMMA 861 — Queen’s University

Pricing the cost of variety in hospital meal planning

A mixed-integer program that feeds patients within clinical limits, on budget — and tells the manager exactly what menu variety costs.

Hospital food managers face a trilemma: cut ingredient cost, satisfy strict clinical nutrition bounds, and keep menus varied enough that patients keep eating. I built a MILP that solves all three at once, then used shadow prices to show which clinical rule drives the most cost.

constraint families encoded
11constraint families encoded
meals per optimization block
9meals per optimization block
patient baseline modelled
400patient baseline modelled
Mixed-Integer ProgrammingPythonSciPy (HiGHS solver)PlotlyShadow Price Analysis
Read the case study
Normal MNormal FDiabeticHigh CholDASH

3-day food cost per profile — cost-only, 3-day rotation, 2-day rotation

Predictive Modelling

Models framed around the decision they enable, not the leaderboard.

Predictive ModellingMMA 860 — Queen’s University

Finding undervalued footballers in Europe’s Big Five

A market-value model over 4,056 player-seasons that separates what a player does from what the market pays for it.

Transfer fees are set by narrative as much as by performance. We merged three sources into one 203-feature dataset, tested four hypotheses about how the market actually prices players, then modelled valuation to surface the gap between output and price.

player-seasons across Europe’s Big Five
4,056player-seasons across Europe’s Big Five
features after merge and engineering
203features after merge and engineering
formal hypotheses tested
4formal hypotheses tested
Pythonpandasscikit-learnstatsmodelsANOVA
Read the case study
KaneWirtzLejeuneMusialaØdegaard

actual vs model-predicted value (€M), top 5 undervalued

Predictive ModellingKaggle competition — team entry

Top 20% on Kaggle’s House Prices — by fixing the encoding, not the model

A disciplined preprocessing strategy beat hyperparameter chasing on 79 mixed-type features.

The Ames housing dataset is 79 features of numeric, ordinal and nominal data with missingness that means three different things. Most of the leaderboard gap comes from treating those distinctions correctly — not from a fancier estimator.

leaderboard finish
Top 20%leaderboard finish
raw features before encoding
79raw features before encoding
CV on RMSE for model selection
5-foldCV on RMSE for model selection
Pythonpandasscikit-learnRidge / LassoGradient Boosting
Read the case study
Top 20%1stlast

Final leaderboard position out of all entries

Predictive ModellingMMA 867 — Queen’s University

RiskIQ: pricing auto insurance risk instead of guessing at it

A claim-likelihood model and a deployable underwriting function — where the simplest, most interpretable model beat every tree ensemble we tried.

Built the EDA and the deployable underwriting tool for a four-person team project: a claim-likelihood model on 10,000 policyholder records, framed around the pricing decision it exists to serve rather than around leaderboard accuracy. Logistic regression, the most interpretable model in the comparison, also won on ROC-AUC — which mattered, because insurance pricing is regulated and "the model said so" is not a defensible answer.

ROC-AUC, best model (logistic regression)
0.887ROC-AUC, best model (logistic regression)
policyholder records, 18 variables
10,000policyholder records, 18 variables
models benchmarked head-to-head
4models benchmarked head-to-head
Pythonscikit-learnLogistic RegressionEnsemble benchmarksClassification metrics
Read the case study
Logistic RegressionLogistic Regression: 0.8865 ROC-AUC0.8865Gradient BoostingGradient Boosting: 0.8814 ROC-AUC0.8814Random ForestRandom Forest: 0.8762 ROC-AUC0.8762Decision TreeDecision Tree: 0.8729 ROC-AUC0.8729

ROC-AUC on the held-out test set — logistic regression beat every tree ensemble

Analytics & BI

Datasets and reporting that other people make decisions on.

Analytics & BICowlar Design Studio (Y Combinator–backed)

A fractional billing system running six figures a month

Analytics-ready SQL datasets and Power BI reporting behind a live production billing pipeline.

At a Y Combinator–backed studio I owned the data layer behind fractional billing — the SQL datasets, the QA process that kept them trustworthy, and the Power BI reporting that turned project financials into something clients and leadership could act on.

billed monthly through the system
$100K+billed monthly through the system
SQLPower BIGoogle Apps ScriptData QA / UATProcess design
Read the case study
Project time& expenseSQL datasetsQA / UATPower BITHE DATA LAYER I OWNED→ client invoicing · leadership reporting

Schematic of the pipeline, not measured data — the billing figures belong to a former employer. marks the QA gate I built.

Agentic AI

Tools I built so the analysis actually acts on itself.

Agentic AIPersonal engineering project

A browser agent that learns each site it visits

Zero-dependency automation that fills any web form — and gets cheaper to run every time it sees a site again.

A universal form-filling agent built on the Chrome DevTools Protocol. A deterministic matcher handles what it can, the language model resolves only what it can’t, and a learning loop turns each run into cached knowledge so repeat visits cost almost nothing.

runtime dependencies
0runtime dependencies
layers of learned site memory
3layers of learned site memory
Node.jsChrome DevTools ProtocolClaude APIAgentic workflows
Read the case study
Agentic AICowlar Design Studio

An HR agent that writes its own paperwork

Text-prompt automation for letters, onboarding checklists and worklogs — built on Sheets, Apps Script and an LLM.

Before “AI agent” was a product category, I built one for HR operations: generate employment and experience letters, produce onboarding checklists, maintain worklogs that feed billing, and read and write the company database — all from plain-language prompts.

Google Apps ScriptLLM APIsGoogle SheetsSlack APIWorkflow automation
Read the case study

Toolkit

What I reach for.

Grouped by what it’s for, not by how long the list is.

Analysis & Modelling

  • Python
  • pandas
  • scikit-learn
  • XGBoost
  • statsmodels
  • PuLP / Solver
  • SQL
  • R

Decision Science

  • Linear & mixed-integer programming
  • Regression & regularization
  • Hypothesis testing / ANOVA
  • A/B testing
  • Forecasting
  • Simulation

Communication

  • Power BI
  • Tableau
  • Looker Studio
  • Advanced Excel
  • Executive decks
  • Data storytelling

Engineering & AI

  • Claude & OpenAI APIs
  • Agentic workflows
  • Node.js
  • REST / GraphQL APIs
  • Google Apps Script
  • n8n
  • Git

The short version

I came into analytics through operations — three years of support and billing work where I kept hitting the same wall: the data existed, but nobody had built anything that wouldact on it. So I started building. A fractional-billing system that now runs real money. An HR agent that writes its own documents. A browser agent that fills in forms by itself.

Now I’m finishing a Master of Management Analytics at Queen’s, which gave me the formal half — optimization, inference, machine learning — to go with the engineering half. In between I started Hopefield, a home-textiles brand, mostly because I wanted a business where I owned every number and couldn’t hide behind someone else’s dashboard.

That’s the thread: I like problems where the answer has to survive contact with a real constraint — a budget, a warehouse, a solver, a customer.

More about me →