CASE STUDY 02APPLIED AI

Connecting business knowledge to daily work

An AI-powered business platform in development across Finance, Marketing and Customer Support.

ROLE / CONTEXTProduct development
PERIOD / STATUSIn development
TOOLS / AREASPython · FastAPI · React · PostgreSQL · Supabase · Prophet

01 / CONTEXT

The problem

A small business can have useful information spread across documents, financial records and customer conversations. Moving between separate tools makes it harder to turn that information into a useful answer or plan.

Opsly is a business platform in development across Finance, Marketing and Customer Support. My work centers on Python backend services, a retrieval-augmented assistant, forecasting and analytics.

02 / PERSONAL SCOPE

My contribution

  • Building Python backend services and a retrieval-augmented assistant for business information.
  • Developing forecasting and analytics across three product areas: Finance, Marketing and Customer Support.

03 / OUTCOME

What changed

Opsly brings Finance, Marketing and Customer Support into one AI-powered business platform, with retrieval and forecasting capabilities in development.

FORECAST WALKTHROUGH · SAMPLE RECORDS

The history changes the forecast.

Compare two example spending histories through Opsly’s local forecasting fallback.

Steady spending: six historical months and three forecast monthsThe solid line shows example records from January through June. The dashed line shows the local fallback forecast from July through September. Exact amounts are available in the data table below.02.5k5k7.5kJanFebMarAprMayJunJulAugSep
Historical recordsForecastEstimated range
NEXT MONTH · JULY$5,160

Estimated range
$4,386 – $5,934

A relatively flat history produces a small change.
View the sample values
Example records and generated forecast values, in US dollars
MonthTypeAmount
JanHistorical record$4,800
FebHistorical record$5,100
MarHistorical record$4,900
AprHistorical record$5,200
MayHistorical record$5,000
JunHistorical record$5,100
JulForecast$5,160
AugForecast$5,220
SepForecast$5,280

THE SYSTEM IN PRACTICE

Knowledge and forecasts, with their own paths.

The platform combines a shared product interface with separate workflows for business knowledge and financial data.

FOLLOW THE FLOW

01 / Workspace

A shared business workspace.

The React frontend organizes Finance, Marketing and Customer Support. Supabase authentication identifies the user making a request.

REACT + SUPABASE AUTH

02 / Knowledge

Retrieve context before answering.

The chatbot backend retrieves relevant document context through embeddings and pgvector search. The assistant can use that business information when responding.

DOCUMENTS → RETRIEVAL → ASSISTANT

03 / Forecast

Turn historical records into a forecast.

The finance workflow groups records by month and produces a forecast from the user’s current history, with different handling when very little data is available.

FINANCIAL RECORDS → MONTHLY SERIES → FORECAST

04 / Review

Bring the result back to the person using it.

Forecast views separate historical and predicted values. Conversation views let the business review assistant sessions and messages.

CHARTS + CONVERSATION HISTORY

ENGINEERING CONSIDERATIONS

The details behind the interface.

Sparse data needs a different response.

The forecasting implementation handles a single month of history with an average-based forecast. With more history it uses Prophet, with fallback behavior when model prediction fails. This makes the amount and quality of available data an explicit part of the workflow.

Keep retrieval behind a service boundary.

The main backend delegates document retrieval to the chatbot service. Embedding and vector-search work stay together, while other product features can request relevant business context through a defined endpoint.

LOOKING AHEAD

Where I’d take it next.

  • Evaluate retrieval against a set of representative questions and expected source passages.
  • Backtest forecasts across different lengths of business history and make fallback behavior clear in the interface.
  • Improve source visibility so a person can check the information behind an assistant response.