Hi, I'm Nishanthan Janarthanarajah · Nishy
I started Agent Studio and Flows to get AI out of the demo and into daily work. I bring that founder's ownership to every client build and every team I lead: 7+ years of full stack engineering, production RAG and agent systems, and engineering leadership as Associate Technical Lead at Eight25Media.
New product · live now
“Describe the process. We run it.”
Flows is the product I founded: an agentic automation platform for small businesses and agencies. You describe a process in plain English, and Claude drafts the workflow, wires up your apps, runs it around the clock, and tells you in one sentence when something needs you. Under the hood it runs on Next.js with Neon serverless Postgres as the backend, so every run and its history is durable and queryable.
New product · live now
“Design the agent. We run it as an API.”
Agent Studio is the second product I founded: a no-code builder for production AI agents. You design an orchestrator and its sub-agents on a canvas, wrap them in guardrails and a prompt-injection shield, test in a playground with a full trace, and hit Deploy. Models run on Groq, and every call, version and trace is stored in Neon serverless Postgres, so ten minutes in you have a live endpoint you can call from any language.
Two apps · one pool of tools
My two live apps, Flows and Agent Studio, and the services, models and infrastructure behind them. Spin the ring, pick anything, and read what it does and where it runs.
My app · live now
“Describe the process. We run it.”
Describe a business process in plain English. Flows connects your apps, builds the workflow, and runs it for you.
Custom builds for clients
Flows and Agent Studio cover the common jobs. For everything else I design and build the AI agent, workflow or assistant your business actually needs, on your data and your tools, for clients in Sri Lanka and worldwide.
Agents that plan, call your tools and finish real business processes, with guardrails and approval gates.
Lead capture, reporting, onboarding and support triage wired across the tools your team already uses.
Assistants over your own data and AI features inside your product, built to production standard.
Key Projects
7 projects across AI agents, B2B platforms, IoT and enterprise data. Pick one, or let the deck play.
Retrieval-Augmented Generation assistant on Azure AI Foundry and Azure OpenAI that answers employee and leadership questions over internal HR data, with Pinecone vector search and role-aware prompt guardrails.
Enterprise RAG HR Assistant
Career
5 roles across 4 companies, from hands-on full stack engineering to leading engineering teams and client delivery.
Oct 2025 – Present
Eight25Media
Engineering Lead / Engineering Manager for a B2B marketing platform, a multi-environment Next.js 15, Contentful and Vercel deployment. Directed a team of engineers through architecture, implementation and production; architected a dual-layer caching system (Next.js "use cache" plus Vercel CDN policies) that resolved a 62.9% cache-miss rate; engineered a multi-tier Contentful–Smartling localization pipeline; drove post-migration technical SEO (JSON-LD, metadata, sitemap, hreflang); and established a dual-remote Git workflow and CI/CD pipeline across the Eight25Media and GitHub organizations while owning client-facing technical delivery.
Intermission
Fun fact: studies estimate the average smartphone user thumbs through about 90 metres of content a day, roughly the height of the Statue of Liberty. Your thumb is doing a triathlon and nobody clapped.
Scrolled on this page
0.0m
0% of a Statue of Liberty (93 m). Keep going, or take a break below.
So here is a 20-second break. Squash a few bugs, beat your best, then keep scrolling. The rest of my work is right below.
Score
0
Streak
–
Time
20s
Best
–
I'm looking to partner with teams that want to build agentic AI workflows and LLM-powered apps — from RAG assistants over their own data to MCP servers that let AI agents do real work
Featured work
Move your cursor over a project to look around it. Each one is a real system that shipped to real users.
01 / 07Featured project
Designed and built an enterprise Retrieval-Augmented Generation (RAG) assistant that answers employee and leadership questions over the company's internal HR knowledge base, including policies, handbooks, org data, and employee records, in natural language.
Vector layer: ingested, cleaned, and chunked internal documents, generated embeddings, and indexed them in a Pinecone vector database with metadata filtering; implemented top-K semantic search with relevance scoring and re-ranking to retrieve the highest-signal context per query.
Generation layer: orchestrated retrieval and generation through Azure AI Foundry, using Azure OpenAI Service GPT chat-completion and text-embedding deployments, Prompt Flow for orchestration, and Foundry evaluation and observability for answer quality, producing grounded, citation-backed responses.
Security & governance: engineered a layered prompt architecture with a base system prompt plus a dedicated security prompt that enforces role-based access control at inference time, so executives receive full organizational data while interns and junior roles are limited to permitted, non-sensitive records.
Responsible AI: grounding and hallucination mitigation with explicit "not found" fallbacks, Azure AI Content Safety filtering, prompt-injection defenses, and PII-aware handling of employee data.
02 / 07Featured project
Built a Model Context Protocol (MCP) server that exposes Figma, Contentstack / Contentful, and a Next.js codebase as callable tools to an LLM agent, turning design handoff into an automated, agent-driven pipeline.
Design ingestion: the agent reads Figma design context (frames, layers, auto-layout, design tokens, typography and spacing variables) and translates it into production-ready React / Next.js components in TypeScript that match the existing component library and design-system conventions.
Data binding: automatically maps each generated component to the corresponding headless CMS content model and wires up live content via GraphQL / Content Delivery API, so generated UI ships bound to real CMS data rather than static markup.
Agent engineering: tool/function calling, schema-constrained structured output, deterministic code templates, and validation passes so agent output compiles and conforms to lint, accessibility, and type-safety standards, with human-in-the-loop review before merge.
03 / 07Featured project
Built a write-capable Model Context Protocol (MCP) server for Contentstack / Contentful that gives an LLM agent authenticated, governed write access to the CMS, automating work previously done manually by content engineers.
Content modelling: the agent designs and provisions content models and content types from natural-language or design-derived requirements, including fields, data types, validations, references, localization settings and relationships, using schema generation and the CMS Management API.
Content automation: generates and publishes entries in bulk, populating new models with structured content and handling references, assets and locale variants, with idempotent writes, dry-run previews, and rollback safety so repeated agent runs never duplicate or corrupt production content.
Governance: human-in-the-loop approval gates, scoped API credentials, and sandbox vs. production environment separation keep autonomous writes auditable and reversible.
04 / 07Featured project
Built an internal AI interviewing and screening application on Azure AI Foundry that conducts and evaluates candidate interviews and automatically shortlists qualified candidates for the hiring team.
CV intelligence: parses and structures candidate CVs with Azure AI Document Intelligence, then semantically matches skills and experience against role requirements using embeddings and vector similarity.
Answer & behaviour analysis: uses Azure OpenAI Service models as an LLM-as-a-judge scorer against a structured competency rubric, with Azure AI Language for sentiment, key-phrase and tone signals, to assess both the substance of answers and behavioural indicators.
Decisioning: combines CV fit, answer quality, and behavioural scores into a weighted, explainable ranking with a per-candidate rationale for recruiters, reducing manual screening effort and standardising evaluation.
Responsible AI: rubric-based consistency checks, bias-mitigation prompt design, Azure AI Content Safety, and mandatory human review before any hiring decision, keeping the system decision-support rather than decision-making.
05 / 07Featured project
Led a cross-functional team of engineers through the end-to-end migration of Paycor's reporting and client data from on-premises infrastructure to Snowflake, owning the project from discovery and architecture planning to final cutover.
Delivered the migration with Snowflake stored procedures and Python automation, coordinated data mapping and schema design, and established validation and testing protocols to catch discrepancies before go-live. The project secured nearly $1M in additional revenue and improved data accessibility, query performance, and scalability for Paycor's reporting and analytics.
06 / 07Featured project
Led the frontend team on Modjoul's IoT asset-management dashboard, translating device telemetry into actionable views for operations teams.
Managed AWS cloud deployments with secure API Gateway and Cognito integrations, and guided a sub-team through the full SDLC.
07 / 07Featured project
Built an online hiring platform that screens candidates using emotion-based analysis and facial recognition, giving recruiters a faster way to shortlist applicants.
Handled both backend logic and frontend UI integration, including cloud deployments.