Work
Selected work
Five engagements where the technical call changed the business outcome: agentic AI in capital markets, a storefront rebuilt for conversion, a comms platform led by a small engineering team, published legal-AI research, and a field-sales suite shipped across Asia.
Case studies
Diagram only — confidential engagement.
Agentic AI for capital-markets infrastructure
Exchanges and data vendors push hundreds of notifications a day; teams triaged that flood by hand. I helped build the agentic workflows that read it autonomously — and the SSO, integrations, and audit trail that let a bank's security team approve the rollout instead of blocking it.
- Enterprise SSO: SAML 2.0 and OpenID Connect, multi-tenant and isolated from the first commit
- Idempotent third-party integrations that hold through retries, partial failures, and bad upstream data
- Explainable, auditable agent output — status tracking, revision flags, a full event timeline
A self-serve storefront scaled to enterprise tier
ServiceBot is WorkWave's online self-serve storefront for field-service businesses. I was the first person from the Sri Lankan team on it and worked across the full product: UX, performance, a new product (Service Express), four suite integrations, fan-out notifications, TypeScript migration, code-level security, and product collaboration with the product owner.
- Service Express shipped: a simplified express storefront as a distinct v2 product
- Integrated with PestPac, Real Green, Service Assistant, and Sales Center: standalone to suite
- Fan-out notifications, ~30x performance gain, TypeScript migration, WordPress plugin
Recreated mockup — confidential engagement.
One inbox, every channel
Communications is the backbone of property management: every maintenance request, renewal, and payment notice runs through it across email, SMS, push, and in-app. I led the small engineering team that owned it at a $35M-funded proptech, building the multi-tenant layer that let new agency partners onboard without custom work.
- Multi-tenant, white-label comms — per-agency branding, sender identities, and schedules
- Kustomer CRM integration; surgical legacy → TypeScript modernization under production load
- Led standups and code review across the team; moderated the cross-team tech-leads meeting
JuriLens — AI that reads case law
Sri Lankan court judgments run past 10,000 words of dense legal language, and the official sites are hard to search. Built as MSc research, JuriLens is a RAG and prompt-engineering pipeline — retrieve, summarize, synthesize, classify — that turns those judgments into structured legal intelligence, with an honest retrospective on what the evaluation got wrong.
- 1,200+ judgments processed across an 89-category taxonomy, built for ~$30 in API spend
- Five LLMs compared — a small open-source model beat GPT-4o on summarization (0.908 vs 0.885)
- Peer-reviewed and published, with the evaluation's limitations stated in full
Recreated mockup — confidential engagement.
SalesSuite — a sales team in your pocket
Field sales ran on paper notebooks, phone calls to HQ, and end-of-day data entry — late orders, blind inventory, day-old numbers. I rebuilt the suite that fixed it: offline-first cross-platform mobile (one codebase, iOS + Android), GPS routing, and an API-first backend. Intern to de-facto architect over four years, because the work landed.
- Native Android → one cross-platform codebase (iOS + Android), designed offline-first for the field
- Legacy PHP/CakePHP → a Node.js, API-first backend; MySQL + MongoDB; cost-tuned infra
- Architected a modular five-product suite shipped across Asia (FMCG, automotive, apparel)
Have a problem that looks like one of these?
Tell me the problem, the constraints, and what ‘fixed’ looks like. If I’m not the right person, I’ll say so.