LLM & RAG Engineering
Retrieval-augmented generation grounded in real knowledge bases: embeddings, semantic search and vector retrieval that keep model output tied to source material instead of invention.
- RAG
- Embeddings
- Vector Search
- OpenAI
- Gemini
Kerala, India
Portfolio / Selected systems
Kasinathan
AI Software Engineer
LLM • RAG • AI Agents • Speech AI • Full-Stack
I build production-grade AI applications and the full-stack systems around them — LLM and RAG pipelines, AI agents and speech-to-text, delivered end to end from interface to cloud deployment.
- Focus
- LLM · RAG · AI Agents
- Experience
- 2+ years, production
- Scale
- 4k–20k monthly users
- Runtimes
- React · Node · .NET
01What I Build
Not features. Complete systems.
Five disciplines that keep collapsing into one workflow — an idea becomes an architecture, the architecture becomes a product, the product becomes infrastructure that runs itself.
AI Agents & Workflow Automation
Multi-agent workflows with function calling and structured JSON output — routing enrichment, qualification and document generation to specialised agents, with prompt and context design tuned for cost per request.
- AI Agents
- Function Calling
- Structured Output
- Context Engineering
Speech & Real-Time AI
Streaming audio capture and Whisper-class transcription wired into live product features — including a clinical pipeline that turns unstructured conversation into structured, section-aware records.
- Whisper
- Speech-to-Text
- Streaming Audio
- Real-Time
Full-Stack Product Engineering
React, Next.js and Angular front-ends over Node/Express and .NET services, with REST contracts, JWT and OAuth 2.0 auth, MongoDB and MySQL data models — shipped as one product, not disconnected layers.
- React
- Next.js
- .NET
- REST
- OAuth 2.0
Automation & Cloud Delivery
Browser and end-to-end automation frameworks in .NET and Python, AI-assisted test generation, and production deployment on AWS with Docker, Redis caching and performance profiling.
- E2E Testing
- .NET
- AWS
- Docker
- Redis
02Selected Work
Project index
Six builds spanning AI SaaS, automation frameworks, identity services, real-time systems and client web work. Open a row for the technical record.
Overview
A browser extension that listens to a doctor–patient consultation, transcribes it in real time, generates a structured SOAP note with an LLM, and writes it straight into the EMR — removing manual note-taking from the clinician's workflow entirely.
Pipeline
Real-time audio capture and streaming → speech-to-text → prompt design that maps unstructured clinical dialogue into Subjective, Objective, Assessment and Plan sections → structured write-back into the EMR.
- 01Capture consultation audio in the browser
- 02Stream to speech-to-text transcription
- 03Map dialogue to SOAP structure via LLM
- 04Validate sections against the record schema
- 05Write the note directly into the EMR
Key capability
- Real-time audio capture and streaming transcription
- LLM-generated SOAP notes in the structure clinicians expect
- Direct EMR write-back — no copy-paste step
- Chrome Extension APIs, inside existing clinical tools
- Prompt design tuned for unstructured clinical dialogue
Stack
- Chrome Extension APIs
- Speech-to-Text
- LLM
- React
- REST APIs
Role
Pipeline architecture · Prompt design · Front-end · Integration
Overview
An AI sales automation platform that centralises inbound leads from WhatsApp, Instagram, Facebook and Google Ads into one CRM pipeline, then qualifies and summarises them with an LLM before a rep ever opens the thread.
Architecture
React client · Node/Express API · MongoDB persistence · Redis caching · Gemini for qualification and conversation summarisation · embeddings and vector similarity for natural-language retrieval across the CRM · deployed on AWS with Meta Graph API ingestion.
Key capability
- LLM lead qualification and conversation summarisation
- Semantic search over leads and message history
- Multi-agent routing: enrichment, qualification, follow-up drafting
- WhatsApp, Instagram, Facebook and Google Ads ingestion
- Lead-source tracking, team assignment and reminders
- Redis caching with MongoDB persistence on AWS
Stack
- Gemini
- Node.js
- React
- MongoDB
- Redis
- AWS
- Meta APIs
Role
Architecture · Full-stack · AI integration · Deployment
Overview
A .NET browser automation testing platform built around a zero-token architecture: repetitive steps replay deterministically with no model calls, and inference is reserved for genuinely novel scenarios — cutting the per-run cost that makes LLM-driven testing impractical at scale.
Approach
A deterministic execution engine handles known paths; an LLM escape hatch handles unseen states. Every resolved scenario is captured, so repeating it later costs nothing.
Key capability
- Deterministic replay of known steps — zero inference
- Model calls reserved for novel or changed scenarios
- AI-assisted test generation from existing flows
- End-to-end browser automation across product modules
- Per-run inference cost measured and driven down
Stack
- C# / .NET
- Browser Automation
- E2E Testing
- LLM
Role
Framework design · Cost architecture · Automation
Overview
A document-driven automation framework that transforms complex requirements into structured, executable rules and validates real-world implementations automatically.
Rule Pipeline
PDF ingestion → document processing → requirement extraction → rule generation → automated validation → reporting. A Python core with C#/.NET service components, exposed through CLI modules and an MCP server for AI-assisted execution.
Key capability
- Document ingestion and parsing
- Requirement extraction into structured rules
- Reusable, configuration-driven validators
- Automated website and API validation
- AI-assisted document understanding
- MCP server for tool-driven execution
Stack
- Python
- C# / .NET
- MCP
- PDF Processing
- LLM
Role
Framework design · Automation architecture · AI workflow
Overview
A high-performance marketing site for an interior design studio: modern responsive UI, on-page SEO, and a self-serve blog management system so the client team can publish without a developer.
Approach
A Next.js front-end tuned for load performance and search visibility, with a content management flow the studio operates themselves.
Key capability
- Responsive marketing site with a modern visual system
- On-page SEO optimisation
- Self-serve blog management for the client team
- Performance-focused page delivery
Stack
- Next.js
- Responsive UI
- SEO
- CMS
Role
Design implementation · Front-end · Performance
Overview
An experimental multiplayer open world built to explore real-time state synchronisation: a large pixel-art forest with a day/night cycle driven by local system time, room-based sessions and live player presence.
Architecture
Socket.io transport with room IDs, usernames and player limits; event-driven flow for movement, survival and combat state; client-side minimap and interaction systems.
Key capability
- Room-based multiplayer sessions
- Real-time player state synchronisation
- Day/night cycle tied to system time
- Minimap, vehicles and world interaction
- Survival, combat and winner resolution
Stack
- JavaScript
- Socket.io
- Canvas
- Real-Time Networking
Role
Game systems · Networking · Interactive UI
03Process
Requirements to optimization
The same seven-stage spine runs through every build. On desktop the pipeline scrubs horizontally as you scroll.
01stage
Requirements
Understand the actual problem, the users, the workflows and the business constraints before a single file exists. Scope, edge cases and success criteria written down.
- User workflows
- Constraints
- Success criteria
02stage
Architecture
Data models, service boundaries, API contracts, auth strategy and infrastructure shape — decided up front so the system can grow without being rewritten.
- Data models
- API contracts
- Auth strategy
03stage
Development
Modular, component-based implementation across front-end and back-end. Small surfaces, explicit configuration, no hidden state.
- React / Node
- Python / .NET
- Modular design
04stage
Integration
The layer where most products break: LLM providers, speech services, third-party APIs, webhooks, message platforms and auth providers — orchestrated with real error handling.
- OpenAI / Gemini
- Whisper / Speech
- Meta / WhatsApp
05stage
Testing
Automated validation over manual repetition: reusable validators, end-to-end browser automation, API checks and AI-assisted test generation executed from configuration.
- E2E automation
- API validation
- AI-assisted checks
06stage
Deployment
Production environments on AWS, Vercel or Firebase — Docker images, environment variables, domains, SSL, storage and caching configured deliberately.
- EC2 / S3 / Redis
- Docker
- Env config
07end
Optimization
Measure, then tighten: render cycles, bundle size, API call patterns, infrastructure spend — and token usage, where prompt and context design decide the cost of every request.
- LLM cost
- Front-end performance
- Caching
04Stack
The toolchain, uncompressed
Everything currently in rotation, grouped the way it actually gets used. Switch namespaces, collapse the drawer, copy what you need.
stack@systems : ~/ai
// intelligence layer — retrieval, agents and grounded generationconst ai = {models: ["OpenAI API", "Google Gemini", "Prompt Engineering", "Context Engineering"],retrieval: ["RAG", "Embeddings", "Semantic Search", "Vector Databases"],agents: ["AI Agents", "Multi-Agent Workflows", "Function Calling", "Structured JSON Output"],applied: ["Conversational AI", "AI Workflow Automation", "LLM Cost Optimization", "MCP servers"],};› 16 entries loadedCategoryCapability
AI & LLMRAG, embeddings, semantic search, vector databases, prompt & context engineering
AI AgentsMulti-agent workflows, function calling, structured JSON output, workflow automation
Speech AIWhisper API, real-time transcription, streaming audio capture, voice integration
FrontendReact, Next.js, Angular, Flutter, browser extensions, Tailwind CSS
BackendNode.js, Express, .NET, C#, Python, REST API design
AuthenticationJWT, OAuth 2.0, Firebase Auth, OTP and verification workflows
DatabasesMongoDB, MySQL, Redis, Firebase, vector databases
Testing & AutomationBrowser automation, E2E, .NET frameworks, AI-assisted test generation
Cloud & DevOpsAWS EC2 & S3, Docker, Vercel, Firebase Hosting, production deployment
Healthcare DomainEMR, clinical documentation, SOAP note generation
PracticesSystem design, scalable API architecture, performance & LLM cost optimization
05Lab
Ongoing experiments
Where emerging technology gets stress-tested before it earns a place in production work.
AI
Context engineering & cost
Tuning prompt design, context construction and token usage so answer quality goes up while the cost of every request goes down.
MCP
MCP developer tooling
Building MCP servers that expose validation frameworks and document pipelines as callable tools, so an LLM can drive real automation instead of describing it.
MOBILE
Story-based AI matching
A cross-platform Flutter concept built on narrative storybook profiles with AI-assisted compatibility matching instead of attribute filtering.
WEB3
Integrity anchoring patterns
Canonical serialisation, Keccak-256 hashing and on-chain anchoring as a general pattern for tamper-evident off-chain records.
AGENTS
Structured output contracts
Function calling and schema-constrained JSON, so model output is something code can consume directly rather than a paragraph that needs parsing and hoping.
SEARCH
Semantic retrieval over records
Embeddings and vector similarity across operational data — asking a CRM a question in plain language instead of assembling filters for it.
VOICE
Streaming transcription latency
How early a live audio stream can be turned into usable text, and how much structure can be imposed on it before the speaker has finished talking.
SURFACES
Products inside other products
Chrome Extension APIs as a delivery surface: shipping features into software you do not own, without asking the user to leave the tool they already work in.
MEDIA
API-driven content platforms
TMDB-backed metadata pipelines, Firebase auth and category-driven routing — how far a consumer media product can be taken on public APIs alone.
INFRA
Cost-aware cloud design
Pay-as-you-go architecture on AWS: instance sizing, Redis caching layers and storage strategy measured against real workload shapes.
IAC
Infrastructure as code
Moving environment setup out of console clicks and into Terraform, so a production environment is something you can read, review and rebuild.
CI
Pipelines for validation
Running automation frameworks on every change instead of on request — Jenkins pipelines that treat document and web validation as continuous, not occasional.
06Experience
Where the work shipped
Two years of production AI and full-stack delivery — healthcare documentation, sales automation, and the automation frameworks underneath them.
01Current
Software Engineer
AOT Technologies
June 2024 — Present
- Built and shipped a browser extension for ambient clinical documentation: it captures consultations in real time, transcribes them, generates structured SOAP notes with an LLM and writes them directly into the EMR.
- Engineered the transcription-to-note pipeline end to end — audio capture and streaming, speech-to-text integration, and prompt design mapping clinical dialogue into Subjective, Objective, Assessment and Plan sections.
- Building a .NET browser automation testing platform with a zero-token architecture that replays repetitive steps deterministically, reserving model inference for novel scenarios.
- Develop enterprise healthcare web applications, including EMR modules used in day-to-day clinical workflows, with React and REST APIs.
- Integrate secure payment gateways into patient-facing flows, handling transaction state, failure recovery and compliance-sensitive data.
- Profile and optimise front-end performance across render cycles, bundle size and API call patterns.
- React
- .NET
- LLM
- Speech-to-Text
- Chrome APIs
02
Freelance AI & Full-Stack Engineer
Self-employed · Remote
- Design and ship custom AI-powered products for startups and small businesses, owning requirements, system design, front-end, back-end, data architecture and deployment.
- Build RAG pipelines using embeddings and vector search to ground LLM responses in client knowledge bases, improving relevance and reducing unsupported output.
- Implement multi-agent workflows with function calling and structured JSON output to automate document generation, lead handling and internal processes.
- Develop speech-to-text pipelines with Whisper-class APIs for real-time, voice-driven product features.
- Operate production services on AWS EC2 and S3 with Redis caching and MongoDB, serving roughly 4,000–20,000 monthly active users.
- RAG
- AI Agents
- Whisper
- AWS
- MongoDB
- Redis
03
Software Engineering Intern
Techfriar
June 2024 — December 2024
- Built and shipped MERN stack application features from requirement through production release.
- Developed reusable React components and integrated backend REST APIs across product modules.
- Diagnosed and resolved production defects, improving stability and front-end rendering performance.
- MongoDB
- Express
- React
- Node.js
Education
Diploma in Computer EngineeringCarmel Polytechnic College2021 — 2024 / CGPA 8.32 / 10
THSLC (Higher Secondary)Govt. THS Haripad2020 — 2021 / Score 98%
07About
Engineering approach
I'm an AI software engineer with 2+ years building production-grade AI applications and the scalable full-stack systems that carry them — LLMs, retrieval-augmented generation, agent workflows and speech-to-text, integrated into React, Node.js and .NET products.
The work spans healthcare, sales automation and Web3, and the applications I have shipped serve roughly 4,000 to 20,000 monthly active users. What connects them is not a technology list — it is taking a product from requirements through architecture, integration, testing and deployment without handing off the hard parts.
- Discipline
- Software engineering · Product engineering
- Frameworks
- Python & .NET automation, MCP tooling
- Interfaces
- React, dashboards, design implementation
- Delivery
- Requirements → deployment → optimization
AI That Ships
Not demos — production pipelines with real users, real error handling and a measured cost per request.
Full-Stack Thinking
Front-end, back-end, database, APIs, authentication and deployment — not a single layer in isolation.
System Thinking
How individual components behave together as one system, under real load and real failure.
Integration Engineering
Connecting model providers, speech services, APIs, platforms and databases into a single reliable workflow.
Cost Awareness
Prompt and context design, deterministic execution paths and caching, so inference spend stays proportionate.
Product Mindset
Users, workflows, scalability, infrastructure cost and business requirements alongside the code.
08Résumé
The one-page version
One page, no filler: the AI and full-stack work, the systems actually shipped, and the numbers behind them. Take it into a hiring conversation or keep it for the shortlist.
PDF/298 KB/Updated August 2026/Kerala, India
- Positioning
- AI Software Engineer — LLM, RAG & AI Agents · Speech AI · Full-Stack
- Experience
- 2+ years shipping production AI across healthcare, sales automation and Web3
- Primary stack
- React · Next.js · Node.js · C# / .NET · Python · MongoDB · Redis · AWS
- Education
- Diploma in Computer Engineering, Carmel Polytechnic College — CGPA 8.32 / 10
09Contact
Kerala, India • IST/--:--:--
Let's buildsomething real.
Use the form for anything with a brief attached — it reaches me faster. For a quick question, email works just as well.