Assistants grounded in your documents (RAG)
Answers with citations from your policies, manuals, contracts or laws. When the documents do not cover a question, the assistant says so instead of guessing.
Proof: KODEKSAI Engineering built on your data, checked by a human
I design and build AI features that answer from your own data, show their sources, and fail honestly. I also use AI agents in my own delivery, with every output reviewed by me before it reaches you.
What I build
What I build with AI
LLM features in existing products
Search, summaries, classification and drafting, inside your app and behind your existing permissions model, instead of in a separate chatbot.
Learn more about LLM features in existing productsDocument and workflow automation
Turning PDFs, emails and forms into structured data and actions, with a human approval step before anything changes.
Learn more about Document and workflow automationIn progress
AI for Shopify operations
A support and operations agent over store policies and the catalogue, with human approval before any change to the store.
Learn more about AI for Shopify operationsHow it works
The AI Sprint one use case, real data, two weeks
Validate one AI use case on your real data in two weeks. You finish with a working prototype and a clear decision, not a slide deck.
01Days 1 to 2
Frame
- Pick one use case with real value
- Define how success is measured
- Confirm data access
02Days 3 to 8
Build
- A working prototype on your data
- Retrieval, citations and guardrails
- A human in the loop from day one
03Days 9 to 10
Measure
- Evaluation against the agreed questions and metrics
- A cost per query estimate
- A plan to production, or a clear "do not build"
What you get
- Go / no-go recommendation
- Working prototype
- Evaluation report
- Cost estimate
- Production roadmap
Scope and price agreed after a call.
Principles
Four rules I do not bend
Grounded
Answers come only from your data, with the sources shown.
Evaluated
Measured before launch, with retrieval and answers checked separately.
Human-gated
A person approves anything that changes data or reaches customers.
Private by design
Sensitive data stays out of analytics and third-party training. On reagiraj.ba, the health risk result never reaches GA4.
Proof
Built, measured and written down
AI in my own delivery
I work as one senior engineer with AI agents. They draft, test and document; I decide and review.
Stack
Used in productionExploring
AI and cloud
- Anthropic Claude API
- OpenAI API (embeddings)
- Claude Code (agentic development)
- Supabase
- Vercel
- Cloudflare
- AWS
- Docker
- Terraform
AI engineering (concepts)
- Retrieval-augmented generation (RAG)
- Vector search: pgvector
- Embeddings: text-embedding-3-small
- Document ingestion and chunking
- Prompt engineering and grounding with citations
- Retrieval evaluation
Exploring
- Python and FastAPI
- Agents and tool use
- Model Context Protocol (MCP) servers
- LangGraph
- Evaluation frameworks (Ragas, DeepEval)
- LLM observability (Langfuse)
- Guardrails
Questions about AI work
Will my data be used to train models?
Not through the way I build. I use the Anthropic and OpenAI APIs under their commercial terms, which state that API inputs and outputs are not used to train their models by default. Your documents and embeddings stay in a database you control, and sensitive data is kept out of analytics.
Do we need a lot of data?
No. A grounded assistant needs the documents that contain the answers, not a training set. To evaluate it, we agree on a set of real questions your team asks, with the answers they expect.
What if the AI gives a wrong answer?
Every answer shows its sources, so a wrong answer can be checked and traced. Before launch I measure how often it happens on your questions, and anything that changes data or reaches customers goes through a human first.
Can it run inside our existing app?
Yes. It can live behind your login and respect your users' permissions, so people only get answers from documents they are allowed to see.
What does it cost to run?
It depends on usage: the number of questions, the length of the documents retrieved and the model used. The AI Sprint ends with a cost per query estimate based on your real data, before you commit to production.