Case Studies

Production AI that shipped and stayed up

Eight engagements, in the order they matter: enterprise retrieval inside a Big 4 firm, a voice agent running a call centre round the clock at a 5,000-person US enterprise, and private on-premise AI for a family office that could not send a single document to anyone else's API.

Every claim below is traceable. Where a client left no written feedback, no rating is claimed.

Expert-Vetted top 1% · 100% Job Success Score · Hong Kong & London

Book a CallRead the Cases
01Enterprise RAG Chatbot for a Big 4 Accounting FirmBig 4 client · fully on-premise0224/7 AI Agent Call Center, 5,000-Employee US Enterprise5,000-employee enterprise · live in production03Super Chain: Private On-Prem AI for a Family OfficeNothing leaves the room04Voice AI Product Build247 hours over three months · 5.0 rating05Multimodal RAG Application246 hours · sustained engagement06Natural-Language Query Layer for Property SearchRated 5.0 · shipped to a live product07N8N2MCP — Open Source130+ GitHub stars · MIT licensed08Fractional CTO: ML Architecture & Roadmap ReviewRated 5.0 · advisory engagement
Case 01Big 4 client · fully on-premise

Enterprise RAG Chatbot for a Big 4 Accounting Firm

The problem

Junior auditors at a Big 4 firm were reading thousands of pages of compliance whitepapers and audit guidelines to answer a single question. Every senior became a search engine, and the same question got asked and re-answered across teams that never saw each other’s work.

What we built

A multimodal retrieval system over the full compliance library, running entirely on the firm’s own infrastructure. OCR and table parsing mean the semi-structured exhibits — the rate tables, the disclosure matrices, the scanned annexes — are searchable rather than silently skipped. Retrieval is tuned for citation rather than summary, so the system returns the passage that supports the answer alongside the answer itself.

Why it holds up

In audit work, an answer you cannot cite is not an answer — it is a liability. That single constraint shaped the whole build: the ranking, the chunking strategy, the refusal behaviour when the library genuinely does not cover the question. A confident guess is worse than no answer at all when the output ends up in a working paper.

RAGOCROn-PremiseLangChain
Case 025,000-employee enterprise · live in production

24/7 AI Agent Call Center, 5,000-Employee US Enterprise

The problem

A 5,000-employee US enterprise needed customer service coverage around the clock without adding headcount to every shift. Overnight and weekend volume was low enough that staffing it properly was hard to justify, and high enough that leaving it unanswered cost real business.

What we built

An AI agent that runs the call centre 24/7: real telephony with barge-in handling so callers can interrupt naturally, appointment booking, and tool-calling into live systems so the agent can look things up and write back rather than just talk.

Why it holds up

Production is a different problem from a demo. This one ships with defined handoff to a human when confidence drops, regression tests over real transcripts before every release, and tracing on every call so a 3am failure is findable the next morning. It is live inside the business, not piloted in a sandbox.

Voice AIAI AgentEnterpriseObservability
Case 03Nothing leaves the room

Super Chain: Private On-Prem AI for a Family Office

The problem

A family office holds the ledger. Fund statements, trust deeds, counterparty terms and tax filings are exactly the documents that cannot be posted to somebody else’s API. That is why most offices of this kind had no AI at all — not because the use case was weak, but because every available product required handing the crown jewels to a third party.

What we built

Super Chain: a fine-tuned LLM with retrieval over financial research, macro data and portfolio holdings, running on the office’s own infrastructure. It is fine-tuned on the domain rather than prompted at it, so it speaks the register the analysts actually use, and every answer is composed with its sources cited so a principal can check the working.

Why it holds up

Nothing leaves the room. The privacy constraint is not a feature bolted on at the end — it decided the model, the deployment and the retrieval architecture from the first design session. That is the difference between an AI system a family office can adopt and one their counsel will never sign off on.

LLM Fine-TuningFinTechOn-PremiseRAG
Case 04247 hours over three months · 5.0 rating

Voice AI Product Build

The problem

A commercial voice product needed to work on real calls, not curated demo calls. The client had already learned that a prototype which impresses in a meeting tells you almost nothing about what happens at volume.

What we built

Voice agent development run as a long hourly engagement rather than a fixed-scope prototype: 247 hours from June to August 2025, iterating against real usage the whole way through.

Why it holds up

The length is the point. A voice product does not fail in the first demo — it fails on the hundredth call, on the accents and the interruptions and the edge cases nobody wrote a test for. Sustained hourly work over months is what lets you find those and fix them. The client left a 5.0 rating: “Cat is a talented Python and AI developer, she has a great attitude, quick learner, and overall fantastic experience!”

Voice AIAI AgentOpenAI APIPython
Case 05246 hours · sustained engagement

Multimodal RAG Application

The problem

A retrieval application over multimodal data, where the source material is part text, part table and part image rather than the clean prose most RAG tutorials assume.

What we built

Each format parsed on its own terms before anything reaches the vector store, rather than flattening everything to text and hoping the embedding model sorts it out. 246 hours of sustained engineering across the ingestion, retrieval and evaluation layers.

Why it holds up

That mix of formats is where naive retrieval quietly fails: the chunker splits a table down the middle, the image carries the number nobody indexed, and the answer comes back confident and wrong. Nothing errors, nothing alerts — you simply get wrong answers delivered with total fluency, which is the worst failure mode a retrieval system has.

RAGMultimodalMachine LearningPython
Case 07130+ GitHub stars · MIT licensed

N8N2MCP — Open Source

The problem

Teams adopting the Model Context Protocol almost always start by writing new tools from scratch — rebuilding integrations they already have running somewhere else.

What we built

N8N2MCP converts an n8n workflow into an MCP server, so automation a team already built becomes a tool that Claude, Cursor or any MCP client can call directly. Paste the workflow, map the credentials, deploy, and the endpoint is live. Open source, MIT licensed.

Why it holds up

The tools are usually already there, sitting in the workflow engine, one adapter away from being callable. 130+ GitHub stars is third-party validation of that argument that nobody paid for — anyone can read the source and check the claim.

Open SourceMCPn8nClaude
Case 08Rated 5.0 · advisory engagement

Fractional CTO: ML Architecture & Roadmap Review

The problem

A team was about to commit budget to a machine-learning roadmap that had never been reviewed by anyone who had shipped one. The expensive mistakes in ML are made in the planning, and they do not surface until two quarters of engineering have been spent on them.

What we built

An advisory engagement rather than a build: reviewing the machine-learning plan, the stack and the sequencing before the team committed to it.

Why it holds up

This is usually the cheapest work we do and the most expensive to skip. Most of the value is in what comes off the roadmap, not what goes on it. Rated 5.0, with the client writing: “Very highly recommend working with Cat. The quality of work, going the extra mile & idea/assistance given were truly outstanding. We will continue to work with Cat on many projects.”

Solution ArchitectureMachine LearningAdvisoryAI Strategy

Your problem is probably on this page.

Retrieval that has to cite its sources, an agent that has to stay up, or AI that cannot send your documents anywhere. Tell us which one and we'll tell you what it takes.

Book a CallRent a Data Scientist