RAG, agents and LLM features
AI development for answers you can check
This is for product teams adding AI to software people rely on: search and Q&A over your documents, AI features integrated into an app you already have, or agents that take actions through your APIs. We build the retrieval, the model calls, the evaluation and the monitoring, and ship them into your stack.
The clearest example is Emblem, an AI due diligence platform for private equity. We built the document intelligence underneath it. It has indexed more than 10 million pages, answers with page-level citations, and scored 100% on the Vectara RAG accuracy benchmark across 3,000 queries.
What we build
RAG and document intelligence
Question answering over PDFs, spreadsheets and transcripts, with OCR, hybrid search and page-level citations.
AI integration in your product
Summaries, extraction, classification, drafting and search, added to the app you already have.
AI agents with tool use
Workflows where the model calls your APIs, queries your data and hands off to a person when it should, with every step traced.
Data extraction
Figures like revenue, EBITDA and margins pulled out of documents and reconciled across sources.
Evaluation
Test sets built from real questions, regression runs before each release and tracing in production, so you see when a change makes answers worse.
How we keep answers grounded
A generic chatbot blends sources and sometimes makes things up. Private equity analysts need answers they can defend in an investment committee, so at Emblem every number had to trace back to the page it came from. This is the approach we built there.
Keep the page attached
Every page is read on its own, charts included, and keeps its file name and page number all the way through. Spreadsheets are parsed sheet by sheet.
Search more than one way
Vector search finds meaning, BM25 finds exact terms, and HyDE searches with a drafted answer. Reciprocal rank fusion merges the results.
Check before answering
A separate model pass scores each retrieved page against the question and drops the weak ones. The answer is written only from pages that passed.
Say when it doesn't know
The citation travels with the page, so an answer can't cite a page it never saw. If the documents don't cover a question, the answer says so.
Models, privacy and running cost
We work with models from Google, OpenAI and Anthropic, and with open-source models when data can't leave your cloud. At Emblem, Gemini 2.5 Flash-Lite reads every page and Gemini 2.5 Flash takes a second pass at the ones with charts. Before launch we estimate what each question or document costs to process, so the monthly model bill is a number you planned for.
What you get
Prototype on your data
A working version on a sample of your documents or workflows, scored against an agreed test set.
Ingestion pipeline
OCR, parsing, metadata and retries, so one broken file never stalls the rest.
Retrieval and answers
Hybrid search, a relevance check and cited answers, behind an API your app calls.
Agent workflows
Tool calls into your systems, with limits, approvals and a person in the loop where it matters.
Evals and tracing
Test datasets, regression runs and production traces in LangSmith or a similar tool.
Production deployment
Running in your cloud, with monitoring for latency, cost and answer quality.
What we've done for clients
Emblem
AI due diligence that cites its sources
We built the document intelligence under Emblem's due diligence platform: page-level OCR with Gemini, hybrid retrieval, a relevance check before every answer, and agent workflows on LangGraph with LangSmith for evals. It has indexed 10M+ pages and scored 100% on the Vectara RAG accuracy benchmark across 3,000 queries, and 95.5% on SpreadsheetBench.
Read the Emblem case studyHexis
Rebuilding the platform behind elite sports nutrition
We used Gemini to add photo, barcode and voice food logging to Hexis, the sports nutrition platform trusted by 40% of the professional peloton at the Tour de France.
Read the Hexis case studyHow a project starts
From a free call to the end of the second week.
A free 30-minute call
Bring an example: a few documents, a workflow or the feature you have in mind. We'll tell you whether AI is a good fit and which approach we'd try first.
A fixed price, or hourly
After the call we send a fixed-price quote for a clear scope, or agree on hourly work at $100 to $150 an hour. A proof of concept on a sample of your data is usually a fixed price. Tuning and new features after that are hourly.
The first two weeks
- We collect a sample of your documents and real questions, and agree on what a good answer looks like. That becomes the first test set.
- We build a thin version end to end and score it against that set.
- You see where it works, where it fails and what it would cost to run.
Frequently asked questions
Which AI models do you work with?
Gemini, OpenAI and Anthropic models, and open-source ones when data has to stay in your cloud. We pick per task on accuracy, speed and cost, and keep the model behind an interface so switching later is a small change.
Do we need RAG?
You need RAG when answers have to come from your own documents and show where they came from. If the job is drafting, classifying or summarizing text you already pass to the model, you probably don't, and the build is simpler.
How do you stop the model making things up?
We answer only from retrieved pages that pass a relevance check, carry each citation with its page, and have the answer say so when the documents don't cover a question. No method makes mistakes impossible, so we measure them with a test set and keep tracking them in production.
Can you add AI to our existing product?
Yes. We work in your codebase and your cloud. At Hexis, for example, we added photo, barcode and voice food logging to a product athletes already used.
Will our data be used to train models?
No. We use paid API plans whose terms exclude training on your data, keep your documents in your own cloud account, and can run open-source models there if nothing may leave it.
What does AI development cost?
We charge $100 to $150 an hour, or a fixed price for a defined scope. Model usage is billed to you by the provider, and we estimate it per question or document before launch. For a quick range, get a rough estimate.