Problem

OpenSearch results needed bias filtering. A paid vendor (Coactive) handled it — expensive and opaque.

Approach

Reverse-engineered the vendor's approach. Built LangChain evaluators to score outputs. Iterated prompts systematically until 94% accuracy was achieved. Deployed on AWS Lambda via Serverless Framework with GraphQL integration.

Outcome

In-house solution matched commercial accuracy at a fraction of the cost. Also built a prompt-iterator Claude skill for future model migrations — reusable by the team.

Stack

Python · LangChain · AWS Lambda · Serverless Framework · GraphQL

Problem

OpenSearch cluster running on x86 — suspected ARM would be cheaper but no data to back it.

Approach

Built a benchmarking script comparing x86 vs ARM on the actual workload. ARM came out ahead. Migrated ECS nodes, updated to ECS-optimised ARM AMIs. Along the way: fixed a 7GB memory issue from an _id-field sorting anti-pattern, migrated CI/CD from Jenkins to GitHub Actions.

Outcome

$11,000/year saved. Currently also leading solo migration from OpenSearch 2 → OpenSearch 3 (derived vector feature saves 40% storage).

Stack

AWS ECS · OpenSearch · CloudFormation · GitHub Actions · Python

Problem

Enterprise clients needed their assessment platforms — SHL, Aon, Aon Gate, Pulsify, HackerRank, HackerEarth, and others — connected to a production hiring system processing thousands of candidates daily for DBS Bank.

Integrations

Owned the full lifecycle for each vendor integration: scoping → technical documentation → implementation → testing → QA. No handoffs — each went from spec to production under one owner.

Analytics Dashboard

Built a client-facing analytics dashboard with modular architecture, real-time data, and per-client customisation.

Internal Tooling

Built an internal ticketing system integrated with Freshdesk for cross-platform sync. Built an internal test platform so the ops team could self-serve integration verification — no engineering needed for routine checks.

Stack

Python · Django · DRF · Celery · Django Channels · REST · SOAP · PostgreSQL

What

Single-handedly managing migration of production OpenSearch infrastructure from version 2 to version 3 at Thomson Reuters.

Approach

Solo deployment of dependent stacks, bootstrapping the new cluster, and validating behaviour across the migration boundary. Using the derived vector feature in OpenSearch 3 to reduce storage footprint.

Outcome

40% storage reduction from derived vectors vs the previous version. Full migration in progress.

Stack

OpenSearch · AWS ECS · CloudFormation · Python · Docker