Mohamed Aslam Abdul Gafoor

1. Automation & Tooling Intern · OCBC (via e2 Power)

   Apr 2026 – present

   Python automation, browser automation and observability for an enterprise technology team, inside a restricted corporate environment.

   • Migrated legacy coordinate- and image-based UI automations to Playwright browser automation, and standardised the Playwright setup across the team’s automation machines, removing a recurring source of breakage whenever an interface changed.
   • Designed and built a shared Python logging module (consistent format, levels and per-job log files) and am rolling it out across the team’s automation scripts in place of ad-hoc console output.
   • Designed and built an automation-health dashboard that discovers job logs, classifies each run as succeeded, failed or stalled from its most recent output, and summarises failures, so problems surface without anyone reading raw logs.
   • Maintain Python/pandas automations over large Excel datasets (100K+ rows), adding input validation so malformed or incomplete inputs fail loudly instead of producing wrong reports.
   • Diagnosed scheduling faults in production automations (unintended repeat runs, jobs that silently stopped running) and restored reliable scheduled execution.
   • Moved the team’s automation code into source control, inventoried and risk-classified existing scripts to prioritise hardening work, and ran a knowledge-sharing session on the new workflow.

2. Founder, full-stack developer · StartupLink

   Apr 2025 – present

   Closed student–founder network. Verified onboarding, role-based access.

3. Co-founder, full-stack developer · LocalLoco

   Sep 2024 – present

   Hyperlocal discovery. Community submission, volunteer moderation, QR redemption.

Internship work is described at the level of engineering approach; system names and implementation details are omitted for confidentiality.

1. Open-source developer tooling

   ROUTR

   • GitHub (opens in a new tab)
   • Case study

   A routing layer for AI coding agents: small “router” skills that match a request to the right workflow and tell the agent what to read, in what order, and what to skip.

   Role

   Author and maintainer

   Problem

   Agents with dozens of installed skills pick the wrong one and burn their context window reading whole files.

   Built

   • 28 router skills with explicit “use when / not for” boundaries, and multi-step route plans capped at three steps with one router active at a time
   • A skill validator that runs in CI on every change, and an eval set of 153 prompts for measuring routing accuracy

   Key decision

   Routing is decided from short skill descriptions alone, so the agent loads only the one full playbook it needs.

   Outcome

   MIT-licensed and installable with one command. The README reports 97–99% routing accuracy in blind model runs against the eval set, with the method published so it can be reproduced.

   Stack

   Python · PowerShell · Shell · GitHub Actions

2. Mobile · offline-first

   Nudgio

   • GitHub (opens in a new tab)
   • Case study

   An Android alarm and reminder app that plays the user’s own media, designed to work with no backend and no Internet permission.

   Role

   Sole developer

   Problem

   An alarm has to fire on time even when the OS kills the app or the device restarts, and it has to fail gracefully when the chosen media file has been moved or deleted.

   Built

   • A Kotlin native core for scheduling, notifications and backup, exposed to a React Native UI through a typed native-module bridge
   • A layered TypeScript codebase with a mock bridge, so UI and service logic are testable without a device

   Key decision

   Register only the single earliest alarm with the OS rather than one per reminder. Recorded with the project’s other design choices in 21 architecture decision records.

   Outcome

   In active development, built against a 23-document spec with a requirements traceability catalogue. Not yet released.

   Stack

   React Native · TypeScript · Kotlin · Android AlarmManager · Room · Jest

3. Full-stack web

   StartupLink

   • Live MVP (opens in a new tab)
   • Case study

   A closed network where students and verified student founders find each other.

   Role

   Founder and sole developer

   Problem

   Trust is the product: founder status has to be verified, and each role may only see its own data.

   Built

   • Server-side founder verification in a Cloud Function, so it cannot be bypassed from the client
   • Role-based access enforced in Firestore security rules rather than in UI components
   • Multi-step onboarding that keeps partial progress across sessions

   Key decision

   Authorisation lives in database rules, not the interface, so a modified client still cannot read another role’s data.

   Outcome

   Working MVP deployed publicly. No user numbers yet.

   Stack

   Next.js · TypeScript · Firebase Auth · Cloud Firestore · Cloud Functions · Tailwind CSS

4. Full-stack web · geospatial

   LocalLoco

   • Live site (opens in a new tab)
   • Case study

   Neighbourhood deal discovery: residents submit offers, volunteers moderate them, and QR redemption tells a business which promotions actually got used.

   Role

   Co-founder and full-stack developer

   Problem

   Location queries that stay fast as data grows, and a moderation queue small enough for a few volunteers to clear.

   Built

   • A Next.js front end and a separate FastAPI service so each deploys independently
   • Location queries in PostGIS instead of application-side filtering
   • Separate submission and moderation stages so unreviewed content is never public, plus QR redemption tracking

   Key decision

   Chose PostGIS early, which avoided a rewrite once real data existed.

   Outcome

   Running on its production domain with authentication in place.

   Stack

   Next.js · TypeScript · FastAPI · PostgreSQL · PostGIS · Supabase

5. Hackathon · data and ML

   DoorLens

   • GitHub (opens in a new tab)
   • Case study

   Nebula X 2026 submission that classifies recorded train-door movements from sensor telemetry, with a tool for testing whether recording assumptions change a result.

   Role

   Developer

   Problem

   A small labelled dataset (110 training cycles) and no ground truth for the test set, so any claim about accuracy had to be honest.

   Built

   • A pipeline that validates uploaded CSVs, segments door cycles and runs a frozen classifier
   • A sensitivity probe that rescales or resamples a cycle and re-runs the same model to show whether the label holds

   Key decision

   Models are frozen, hashed artefacts: uploads never retrain them, and every result carries the identity of the model that produced it.

   Outcome

   Submitted with a written limitations document and verification notes.

   Stack

   Python · NumPy · pandas · FastAPI · Next.js · pytest · Docker

Grouped by depth of use rather than self-rated mastery. Strong means used repeatedly in the internship or across several of the projects above.