GenAI document assistant — enterprise RAG
Production
Executives at Emerson needed answers out of internal documentation that lived
across roughly twenty SharePoint documents, and nobody was going to read them.
I ran this one end to end — requirements with the executives themselves, design
with the developers, project documentation and stakeholder coordination, plus
build. Documents are chunked and embedded into pgvector, and answers are generated
from the retrieved context through the Gemini API. The engineering problem was
never the model: it was chunking that preserved meaning, retrieval that returned
the right passages, and grounding answers so they could be trusted by people
making decisions on them.
RAGpgvector
EmbeddingsGemini API
SharePointPython
Report pipeline rebuild — field-service platform
Production
Bulk PDF generation was a platform-stability problem, not a performance one:
10 MB+ inspection reports with HD site photography took over 30 minutes,
repeatedly crashed the server, and blocking calls cascaded into unrelated API
failures. I re-architected it as a dedicated microservice with a MongoDB-backed
job queue processed sequentially by cron, with bounded retries and extended
timeouts on downstream calls. The crashes stopped, and it has since sustained
2,000+ reports per six-month cycle through financial year-end peaks.
Node.jsTypeScript
MongoDBQueue design
Microservices
Platform-wide auth & RBAC rollout
Production
The platform had no role-based access control — every API endpoint was open
to any authenticated caller. I designed and rolled out JWT authentication and
RBAC across the entire API surface, integrated with Okta as the enterprise
identity provider, and built the user-lifecycle flows around it: app-access
requests routed to managers for approval, and staged inactivity warnings at
80/85/90 days with automatic deactivation, backed by MongoDB time-series
collections that also became the source for engagement analytics.
Okta SSOJWT
RBACOWASP
MongoDB time-series
PHP → Node.js migration at IoT scale
Production
Core contributor on a CTO-led migration of a legacy PHP vehicle-tracking
system to Node.js and Express microservices. The PHP batch logic could not
sustain the volume — 10,000+ vehicles each transmitting location and device
data every minute, around 14 million data points a day. I rebuilt the
server-side processing and reporting layer on top of it, deployed on AWS
EC2 with S3 for asset storage.
Node.jsExpress
MongoDBAWS EC2 / S3
IoT telemetry
SmartDocs — Document Intelligence API
In development
A REST API for document upload, text extraction and LLM-backed Q&A over
document content — built to production standards rather than demo standards:
JWT auth, rate limiting, Docker containerisation, GitHub Actions CI. Currently
extending it with semantic search over pgvector, an async job-processing
pipeline, and Prometheus metrics.
Node.jsTypeScript
ExpressMongoDB
pgvectorDocker
GitHub Actions
Agentic finance prototypes
In development
Multi-step agent prototypes on finance workflows — the current one works a
loan file the way an analyst would: pull the documents, reconcile the numbers,
flag what doesn't add up. Built mainly to find out where agentic workflows
genuinely beat a deterministic pipeline, and where they just add latency and
failure modes.
PythonFastAPI
LangGraphLLM APIs
Evals