Alec Bhamani

Data science student at UW–Madison

I build predictive models, machine-learning systems, and practical automation tools. I’m especially interested in work that connects technical analysis to a real decision.

View projects Email me
Alec Bhamani
Open to 2026–27 data and ML internships
UW–Madison B.S. Data Science
1st place Predictive modeling
5 selected projects ML, product & retrieval
May 2028 Expected graduation

Selected work

Selected projects

These projects cover retail prediction, language models, agent tooling, personal finance, and course retrieval. The technical details below come from the work itself.

01 Predictive modeling · Team of four

Crystal Farms & Cub Foods promotion model

Finding promotional windows where discounts could improve profitability—not just lift unit sales.

My contribution
Collaborated on the Python modeling workflow and the analysis of historical retail sales data.
Approach
Converted historical sales patterns into predicted promotional windows the team could compare and present.
Result
First-place project at the UW–Madison Tech Exploration Lab.
PythonPredictive modelingSales analytics
View résumé evidence
02 Language models · Independent build

Decoder-only GPT training stack

A custom PyTorch GPT-2 training stack covering data preparation, pretraining, chat tuning, evaluation, and export.

Build
Implemented a decoder-only transformer with causal attention, tied embeddings, checkpoint recovery, and KV-cache generation.
Engineering
Added GPT-2 tokenization, resumable pretraining, assistant-only chat tuning, perplexity evaluation, and a Hugging Face export path.
Verified scale
The checked-in configurations calculate to 7.27M, 51.24M, and 124.48M parameters.
PyTorchCausal attentionKV cacheNLP
Explore the code
Repository workflow README
  1. prepare_data.py Local text or FineWeb-Edu → GPT-2 BPE → train.bin and val.bin
  2. train.py Resumable decoder-only pretraining → gpt_model.pt
  3. finetune_chat.py JSONL conversations with assistant-token-only loss → gpt_chat.pt
  4. eval.py Perplexity evaluation, then Hugging Face and Ollama-compatible export tooling

Presets calculated from model.py

PresetLayers × widthHeadsContextParams
tiny4 × 12842567.27M
small8 × 512851251.24M
gpt212 × 768121024124.48M
03 Agent tooling · Independent build

skillmap

A Python CLI that indexes installed agent skills and returns only the tools relevant to the current work context.

Problem
As skill catalogs grow, overlapping descriptions make it harder for an agent to select the right capability.
Approach
Discover skills, extract stable concepts, build a graph, and rank a scoped neighborhood for each work query.
Verification
Added redirection-aware ranking, incremental updates, diagnostics, and a 55-test regression suite.
PythonKnowledge graphsRetrievalCLI
Explore the code
A real scoping run Diagnostic source

skillmap scope "manage worktrees and terminals from the command line"

orca-cli Matched on: terminal, worktree

1 of 5 installed skills surfaced

55 / 55
tests passed
8 / 8
curated queries ranked first
3 / 3
adversarial queries ranked first
1.8 / 5
skills surfaced per query

Fresh local diagnostic against current main · 5 installed skills · July 29, 2026

04 Personal finance · Work in progress

Pace

A paycheck-first budgeting app that calculates how much to spend, save, and invest from variable income.

Build
Developing a full-stack TypeScript app with a paycheck-planning engine and D1-compatible persistence.
Logic
Converts pay frequency and monthly category limits into per-paycheck spending, savings, and investable amounts.
Integration
Supports Plaid Sandbox account linking with encrypted access-token handling on the server.
TypeScriptReactD1Plaid
Explore the active build
Paycheck plan calculation Implementation
Typical take-home paycheck$2,000.00
Per-paycheck spending cap− $1,200.00
Savings contribution− $200.00
Available to invest$600.00
$1,200Spend
$200Save
$600Invest

Illustrative Sandbox values · calculation mirrors the current planning engine in app/page.tsx

05 MadData 2025 · Team hackathon

BadgerBot course recommender

A natural-language course search prototype for exploring UW–Madison classes by topic, credits, and completed prerequisites.

My contribution
Built the initial Flask interface and project setup as part of the team; the current retrieval implementation reflects contributions across the repository.
Approach
Embeds 380 catalog records with MiniLM, retrieves candidates with FAISS, and filters by credits and prerequisite gaps.
Output
Returns up to five course cards with course details, credits, and missing prerequisites.
PythonFAISSMiniLMFlask
Explore the team repository
Prototype recommendation Retrieval source
Query

3 credit course about data visualization

Completed: COMP SCI 220, COMP SCI 320
  1. COMP SCI 565
    Introduction to Data Visualization3 credits · Missing prerequisites: None
  2. COMP SCI 765
    Data Visualization3 credits · Missing prerequisites: None
  3. STAT 436
    Statistical Data Visualization3 credits · Missing prerequisite: STAT 240

Prototype output regenerated from the committed FAISS index; ranking percentages intentionally omitted.

Additional product experiments and coursework are available on GitHub.

Browse all repositories

Working toolkit

Tools and methods

Tools matter most in context. These are the methods I’ve used across projects, coursework, and consulting work.

01

Model development

Python, PyTorch, regression, clustering, model evaluation, and experiment design.

02

Data & retrieval

Pandas, NumPy, R, FAISS, statistical analysis, feature preparation, and ranking.

03

Product delivery

Flask, APIs, Git, Java, workflow automation, technical documentation, and team collaboration.

Experience

Experience and education

My background spans consulting, service leadership, and project delivery—the settings that taught me to make technical work clear and usable.

2025—Now

Associate Consultant · Krahnborn

Contribute to data-driven solutions and automation strategies, using AI tooling and analysis to reduce manual work and improve deliverable accuracy.

2024—2025

Bar Manager & Catering Staff · Wisconsin Union

Promoted into leadership and coordinated staff for high-volume university venues and events.

2020—2022

Jr. Project Engineer · Blue Lynk

Delivered 20+ interactive document projects for public and private clients while meeting cloud-based quality standards.

Let’s connect

Get in touch

I’m open to data science, machine-learning, and analytics internship conversations for 2026–27.

Email Alec View résumé