Taha Zubair

Data Analyst

Based in: Lahore, PakistanCurrently: AI Workflow Consultant, FreelanceExperience: 2 years

Experience

AI Workflow Consultant

Freelancework2025 – Present

Summary

Built tailored Claude setups for 3 clients, including brand-guideline specs that generate on-brand decks on command, cutting deck turnaround from hours to minutes. Shipped scheduled agents that send recurring team-completion reports, plus writing and analysis bots tuned to each client's voice and market position, several still running after handover.

What I did

Developed custom Claude Code skills triggered by specific keywords to execute personalized, multi-step tasks automatically.

Results

Built tailored Claude setups for 3 clients using global CLAUDE.md files to enforce brand guidelines, cutting deck turnaround from hours to minutes.
Automated daily reporting workflows using Agent SDKs for Claude Code, ensuring consistent logic and market positioning across client deliverables.
Shipped scheduled agents and analysis bots that remain operational post-handover, maintaining the client's unique voice.
Accelerated project execution for Government of Punjab stakeholders from weeks to days by automating analysis structures for director-level updates.
Engineered keyword-triggered personalized Claude skills within CLAUDE.md to automate specific workflows on command.
Reduced deck turnaround time from hours to minutes for 3 clients.
Successfully deployed automated agents and bots that remain operational post-handover.
Improved execution speed from weeks to days for government-level project updates.

ClaudeAI AgentsRAGClaude CodeAgent SDKs

Data Analyst

Baked (F&B group, Lahore)LahoreworkAug 2025 – Nov 2025

Summary

Built an internal n8n chatbot answering staff questions from 50+ product, policy and compliance documents, handling 200+ queries and keeping an estimated 50% from reaching upper management. Consolidated customer review forms, end-of-period stock counts and in-branch category comparisons gathered across all branches into a Power BI dashboard, then defined 7 KPIs that scored 276 vendors and flagged 46 long-term underperformers for review. Layered ABC-XYZ classification onto the vendor scoring system to sort suppliers into tiers and match support to potential, cutting 5 vendors after repeated issues and growing mid-tier vendor sales 20% through shelf-space moves and in-store pop-ups.

Results

Built a RAG pipeline handling 40 queries daily, reducing management escalations by 50%.
Defined 7 KPIs to score 276 vendors and identify 46 underperformers.
Grew mid-tier vendor sales by 20% through strategic shelf-space moves and pop-ups.
Developed a large-scale Supply Chain Forecasting Model in Excel (~8.9MB) to optimize inventory planning.
Built an interactive HTML POS Sales Analysis dashboard for real-time operational insights.
40 queries/day via RAG-based n8n chatbot, reducing management escalation by 50%.
Cut 5 underperforming vendors after repeated issues.

n8nPower BIABC-XYZ classificationKPIsRAGSupabaseExcel Solver@Risk

Business Development Intern

BakedworkJul 2025 – Aug 2025

Summary

Automated the inventory system in Google Apps Script so daily POS data triggered stock-level checks and pushed replenishment orders to vendors in real time, cutting weekly stock-outs to 6% and lifting order accuracy 11%. Redesigned the operational data collection process, restructuring 4 vendor forms and 25+ free-text fields into standardised, validated inputs, which cut the cleaning work before each reporting cycle by roughly 60%. Rebuilt weekly reporting as a Google Apps Script pipeline (Forms → Sheets → Power BI), replacing 100+ manual entries a week and the transcription errors that came with them.

Results

Cut weekly stock-outs to 6% and lifted order accuracy by 11%.
Reduced data cleaning work by approximately 60%.
Replaced 100+ manual entries per week by automating the reporting pipeline.

Google Apps ScriptGoogle FormsGoogle SheetsPower BIPOS data

Skills

Technical

Google Sheets

Advanced Excel

Vibecoding

SQL

PostgreSQL

Power BI

n8n

Google Apps Script

RAG workflow build

Figma

Claude Code

UI/UX Design

2D Rendering

p5.js

three.js

ggplot2

Machine Learning

Claude Cowork

dplyr

Blender

rpart

cluster

KNIME

k-Means

Next.js

PCA

Islamic Finance Jurisprudence

TypeScript

Projects

Startup Funding Data Quality Monitor

Data Quality Engineeracademic

Designed a PostgreSQL data-quality suite over a startup funding dataset (companies, rounds, investors), with automated checks for completeness, freshness, duplicates, and referential integrity across ~50K records. Wrote 30+ assertion queries using CTEs and window functions to catch null critical fields, stale records past an SLA threshold, fuzzy-duplicate company entities, and round amounts falling outside plausible ranges. Logged every failed check to an audit table feeding a domain-level coverage scorecard, cutting the records failing at least one quality check from 12% to under 2% across three cleanup passes.

SQLPostgreSQLCTEsWindow Functions

Automated Data-Quality Audit Loop

Developerpersonal

Built an agent loop in Claude Code that walks each data domain in turn, runs a defined check suite, and retries or escalates on failure, so a full audit runs from one command rather than a manual pass. Specified the checks in CLAUDE.md as reusable skills, so adding a domain means writing its rules instead of editing the runner, and every run applies the same standard. Wrote each run to a dated markdown summary and an append-only log, turning coverage and failure counts into a week-over-week trend rather than a snapshot.

Claude CodeMarkdownCLAUDE.md

Report Quality Scoring on the NUFORC Dataset

Data Scientistpersonal

Engineered 40 features across 125,290 NUFORC sighting reports in R, scoring each record on data sufficiency, internal consistency and red-flag indicators using AARO's FY24 evaluation criteria as the framework. Clustered the scored reports with k-means, using elbow and silhouette diagnostics to settle on three tiers, which separated the 9.4% of records too sparse to analyse from the 19.7% detailed enough to warrant investigation. Fitted an rpart decision tree to make the tiers explainable, showing that internal consistency, text coherence and red-flag score alone account for 80% of the classification.

Rk-meansrpartFeature Engineeringggplot2dplyr2D Renderingp5.jsthree.js

Vibe-Coded Vintage Literature Library ("shelf")

Creative Lead & Developerpersonal

A full-stack digital bookshelf and reading-progress app (deployed as the Vintage Literature Library). Features a custom 'ShelfEngine' for library management and 'FallingWords' for interactive physics-driven text animations. Implements a unique 'book chooses the reader' recommendation system where users answer questions based on story scenes; recommendations are made if the user's decisions align with the protagonist's. Built with a focus on 'vibecoding' to bridge the gap between intricate design and functional code.

Next.jsTypeScriptShelfEngineFallingWordsFigmaBlenderPhysics Libraries

UFO Sighting Prediction Model

personal

Developed a machine learning model to predict UFO sightings based on historical classification data using scikit-learn, pandas, and numpy for a Business Analytics course project.

Pythonscikit-learnpandasnumpyOpenCV2D RenderingMatplotlibp5.jsthree.js

Google Maps Lead-Gen Pipeline

professional

Orchestrated a lead-generation pipeline to scrape Google Maps business listings and extract contact details. Self-hosted n8n via WSL2/Ubuntu to resolve node-gyp build issues.

n8nApifyFirecrawlWSL2Ubuntu

Bedroom Wall Projection Mapping

personal

Physical installation project projecting physics-driven particle animations across a 10x6ft bedroom wall, mapping over furniture and instruments to transform the physical space.

Map ClubPhysics-driven animations

Islamic Fintech & Shariah-AI Concept

personal

Explored AI-powered angel investor–SME matching platforms with Shariah screening and RAG-based fatwa support. Researched Takaful alternatives and Islamic VC fund structures.

RAGAI Agents

MIDI Classical Music Clustering

academic

Unsupervised ML project analyzing 4,800 MIDI files. Extracted features using Python and built a KNIME workflow for PCA and k-Means clustering. Achieved an Adjusted Rand Index (ARI) of 0.1859, recovering ~19% of true musicological era structure from raw data.

KNIMEPythonPCAk-Means

Education

BSc (Hons)

Lahore University of Management Sciences (LUMS)Management ScienceSep 2022 – Jun 2026

Coursework

Business Data ManagementData Analytics for New Product DevelopmentOptimization MethodsDecision AnalysisProbability & StatisticsOperations Management

Recognition

Certifications

Anthropic Claude Certification - Level 1

Anthropic

Q&A

In Taha’s own words.

How do you handle AI's creative limitations?

When AI tools fail to capture the 'compassion' or intricacy of a vision—like the book covers for my library project—I step in manually. I use Figma, Blender, and curated inspiration to ensure the final output maintains its soul and aesthetic integrity. AI is the partner, but I am the architect.

What kind of data projects actually excite you?

I'm driven by 'meaningful mysteries.' Instead of standard cohort segmentations, I want to dive into environmental data, government systems, or unique creative challenges where the data tells a story that needs a technologist's eye to uncover.

What drives the work

Motivations: Meaningful mysteries (Environment, Government, Social Impact)
Creative expression through code
Building tools that provide access and record
Objectives: Transition into Creative Technologist roles
Work on high-impact social/environmental data projects
Relocate to Europe (Amsterdam/Berlin) or the US
Principles: Creative Technology
Meaningful Data Experiences
Continuous Learning
Art-AI Intersection

Hire TahaReach out about a role, a contract or a conversation.For recruiters

Taha's twin is AI, it can make mistakes.

More people on Kanvis

Browse more people Create your own Kanvis