K M Rakeen Abtahee — Full-stack engineer with 2+ years delivering production software embedded in client environments across enterprise and startup contexts. Owns work end to end: technical design, direct stakeholder communication, and shipped launches. Core stack is TypeScript, React and Next.js on the front end with Python and FastAPI on the back end, alongside production applications built on LLM APIs, RAG systems, agentic workflows, evaluation tooling and vision-based document extraction. Comfortable working inside unfamiliar codebases and platforms without a ramp-up period.
Most of my commit history lives in Bitbucket through work.
How the work arrives
Two years inside client environments rather than on an internal product team. What each one was built with runs down the left.
Three of six client engagements, the three I can describe without naming anyone. New business, new codebase, meter already running.
Mobility & vehicle services
1.5M records migrated
Python · FastAPI · Next.js · Zoho CRM · SQL
National health insurer
Manual entry down 80%+
Python · FastAPI · Computer vision · AWS
Financial services
130+ tickets resolved
Apex · LWC · SOQL · Workato
Nobody assigned these. I noticed the gap, scoped it, built it, and got it adopted, which is a different muscle from delivering a brief well.
PromptBench
30+ users · 2hrs → <30min
TypeScript · Node.js · MCP · Claude API
Expertise Finder
Answers inferred across documents
Python · FastAPI · RAG · Claude API
QA generation
AWS Lambda · Bitbucket Pipelines · LLM
Selected work
Client code isn't public so these are the projects where you can read the source and ask me why I did it that way.
Expo · React Native · TypeScript · SQLite · Claude Vision
Cross-platform mobile app that photographs receipts, extracts and classifies line items via a vision model, and categorises expenses against ATO deduction rules for work-from-home and work-related claims. Offline-first local storage so capture works without connectivity, with export to CSV/PDF at tax time.
Not published to any store, I sideload it onto my own phone and use it. That is the point of it.
Flutter · YOLOv8 · TensorFlow Lite
Mobile real-time object detection app; handles camera frame preprocessing, tensor conversion, and bounding box rendering. Model optimised for edge devices using TFLite.
No network round-trip, so the whole latency budget belongs to the device.
Claude Code · Bash · YAML
Seven-subagent orchestration template (architect, researcher, implementer, debugger, reviewer, security-auditor, test-runner) automating a spec → implement → review → test workflow.
Lifecycle hook scripts and a written CLAUDE.md orchestration contract enforcing review gates. Reusable across projects.
Canvas 2D, then Three.js · custom GLSL shaders
Animated jellyfish aquarium, Canvas 2D, then rebuilt in Three.js with custom GLSL shaders. Built purely to understand shaders.
Built purely to understand shaders. No practical purpose, which was the point, the field further up this page is where that curiosity ended up.
Scala
A Clojure-inspired interpreter written from scratch, custom parser, evaluator and REPL.
Ported to TypeScript and running further up this page, go and type into it.
Python · Claude API · Google News RSS
Sentiment-adjusted Monte Carlo portfolio simulator, overlays live news sentiment (Claude API classifier over Google News RSS) onto Monte Carlo simulations to show how outcome distributions shift.
The interesting part is joining a qualitative signal to a quantitative model without pretending the join is precise.
Next.js · Node.js
Full-stack note-taking platform with authentication, search and a clean UI.
Play with it
Originally in Scala, ported here so you can run it. Three stages — tokenise, parse, evaluate, and a scope chain that makes closures behave properly. Type into it.
Play with it
No scene, no geometry, no engine, one fragment program deciding every pixel independently, sixty times a second. I wrote it to understand how the coordinate space folds, then kept turning the numbers up. Drag your pointer through it and take the sliders; they write straight into the running program.
Raw WebGL2. The field is warped through itself three times, twisted harder toward the centre, then sampled at three slightly different scales so the colour channels come apart, which is where the fringing comes from. Nothing here is a texture.
Experience
Embedded delivery across six client engagements, live systems, other people's deadlines.
Ingenuity Partners
Salesforce and full-stack consultancy; engineers embed directly with client teams.
IBM
Macquarie University
Double major: Software Technology and Data Science
What I build
Three areas, one habit: code that has to survive real users, real data volume and a release date.
React, Next.js and React Native front ends over Python and Node services, deployed on AWS with CI/CD I own myself.
Client SLAs, fixed sprint cadenceRetrieval pipelines, agentic workflows, evaluation harnesses and vision-based document extraction, built where they remove work, not where they demo well.
Evaluated against fixed test suitesEvent-driven backend services in Apex and reactive UI on the native Web Components standard, delivered into live client workflows.
Hard CPU, heap and query budgetsStack
Bold is where I work daily and would happily be interviewed in depth. The rest is real, project-level exposure.
The constraint run
Six snippets. Each has exactly one thing wrong with it. Click the line. All six are drawn from constraints I've actually worked under, a query budget, a live migration, a document corpus somebody else can write to.
Click the line you'd flag in review.
The first snippet above was a query inside a loop. This is what that costs. Salesforce gives every transaction a hard ceiling, 100 queries, 10 seconds of CPU, and going over doesn't slow things down, it throws and rolls the transaction back. Drag the slider.
for (r : records) { SELECT … }
Within budget. Holds until 100 records.
SELECT … WHERE Id IN :ids
Flat. One query at 50 records, same as at 1.
Note, limits are the real synchronous Apex ceilings (100 SOQL, 10,000ms CPU). Timings are illustrative, not measured.
Happy to walk through any of the above, the architecture decisions, the parts that went wrong, or the constraints that shaped them.
rakeenabtahee@gmail.com