Selected work

Real problems we worked on

Examples of what we can build for your company: the first job Attera is building, and four projects we did in collaboration with Albert School.

A thick client file with tab dividers on a deskBanking · KYC

Attera's own build In development

Corporate KYC files, with every answer tied to its page

For every corporate client, a bank reads a KYC file of around 200 pages by hand to fill in a fixed checklist. We are building the assistant that fills it in from the bank's own documents, on a local AI model, with every answer tied to the page it came from and anything missing marked as missing, never made up.

Before200 pages

Read by hand for every corporate file, to fill in one checklist.

After, on public files98 / 116

Fields correct, 0 wrong, 18 marked not found, on 26 public UK company filings (not bank files).

How it worksStep 1 of 4
1

Read the file

Every page, scanned or digital, read and kept with its source.

2

Find each answer

The file is searched for every field on the bank's checklist.

3

Prove it or refuse

Each answer is checked against the line it came from. No line, no answer.

4

A person decides

Answers several documents agree on are marked confirmed. The reviewer makes every call.

Built with
  • Python
  • FastAPI
  • SQL (SQLite)
  • Local LLM (no outside AI service)
  • OCR (DocTR)
  • PyMuPDF
  • Hybrid search (BM25 + vectors)
  • Citation checks
  • Audit log (SHA-256)
Where this fits you

Any team that fills in a checklist from a stack of documents, or needs insights from long documents.

Office towers in a financial district at duskAsset management

(In collaboration with Albert School)

Twenty competitor outlooks, read and compared automatically

Every year, the large asset managers publish their view of the year ahead. Sales teams in five countries need to know where each competitor stands against the bank's own view. We built an AI pipeline that reads them all and does the comparison.

Before20 by hand

Outlooks read one by one, compared market by market.

After20 × 14 × 5

Competitors, forecasts and markets compared in one run, with ready answers for the sales team.

How we got thereStep 1 of 4
1

Collect

The pipeline gathers the outlooks of 20 large asset managers.

2

Read and extract

An AI model pulls the same 14 forecasts out of each document.

3

Compare

Each forecast against the house view, for five markets.

4

Hand to sales

One dashboard and ready answers for every major competitor.

Built with
  • Python
  • Claude API (Anthropic)
  • Web scraping
  • Document parsing
  • Streamlit
  • HTML dashboard
  • CSV and Word exports
Where this fits you

Any team that reads a stack of documents to fill in the same table, again and again.

A boardroom table with printed reportsPrivate banking

(In collaboration with Albert School)

Which families first? A prospect list of 300, scored and ranked

A private bank wanted to win clients among Switzerland's wealthiest families, but had no clear order of approach. We profiled them, scored each one, and gave the bankers a tool to decide for themselves.

Before300

Families on a list, with no order and no way in.

After2 segments

Ranked prospects, each segment with its own approach, in one dashboard.

How we got thereStep 1 of 4
1

Build the list

300 families profiled by age, wealth, sector and region.

2

Score

Six factors each, from wealth to how easy they are to approach.

3

Pick the segments

Industrialists to approach directly, dynastic families to win slowly.

4

Hand over a tool

A dashboard the bankers use without an analyst beside them.

Built with
  • Excel
  • Power BI
  • Scoring model
  • Segmentation
  • Network mapping
Where this fits you

Any team that has to decide, from a long list, who to call first and how.

A quiet office with shelves of bindersPrivate banking, AI

(In collaboration with Albert School)

Giving relationship managers their week back

The relationship managers studied spent most of their week on research, preparation, compliance and admin. We designed AI assistants for those tasks, so relationship managers can spend more of their time with clients.

Before25%

Of the week spent with clients.

After, projected51.5%

With five assistants, paying for themselves by month 15 of implementation.

How we got thereStep 1 of 4
1

Map the week

Where the hours of a relationship manager actually go.

2

Five assistants

Client research, portfolio, compliance, meetings and market news.

3

Build a prototype

A working version with realistic client profiles.

4

Make the case

Costs and savings worked out bottom-up, phase by phase.

Built with
  • React
  • AI agent design
  • Token economics
  • ROI model (Excel)
  • EU AI Act risk classification
Where this fits you

Any team whose best people spend most of the week preparing, checking and filing.

A collector's room with a classic car and paintingsWealth tech

(In collaboration with Albert School)

Only the news that touches what a family owns

A family office follows art, cars, property and more, and the daily news is endless. We built an alert engine on a local AI model that keeps only what concerns the family's own holdings.

BeforeHundreds

Of news items a day, almost none of them relevant.

After90%+

Precision and recall on 31 test cases, with no data leaving the office.

How we got thereStep 1 of 4
1

Read the news

Every morning's articles on art, cars, property and markets.

2

Check the holdings

A local AI model drops everything that does not concern the family.

3

Route it

What remains goes to the family member it concerns.

4

Keep it inside

It all runs within the family office.

Built with
  • Python
  • Local LLM (Qwen 2.5, LM Studio)
  • Embeddings (BAAI bge-m3)
  • ChromaDB
  • Neo4j
  • HDBSCAN
  • spaCy
  • Streamlit
Where this fits you

Any team that has to watch a flood of information for the few items that matter, without sending its data outside.

Our toolkit

The tools behind the work, picked for each job

Data and analysis

Python (pandas, scikit-learn, LightGBM, statsmodels), SQL, R, Excel modelling

AI and language models

Local LLMs (Qwen, gpt-oss) via LM Studio and Ollama, Claude API, embeddings and reranking, RAG pipelines

Data stores

SQLite, ChromaDB (vector search), Neo4j (knowledge graphs), hybrid BM25 search

Interfaces and reporting

Power BI, Streamlit, React, FastAPI, HTML dashboards

About these projects. Each case shows a kind of process we can build for your company. The KYC assistant is Attera’s own build and still in development; its figures come from public company filings, not a bank’s files. The other four were done in collaboration with Albert School, in teams, for the company that set the brief. Figures marked projected are estimates from the project.

Your process will not look exactly like any of these. That is the point: we start from how your team works today and build the version that fits it.

Have a problem like these?

Talk to us