Finance systems · multi-agent architecture · embedded hardware
I build systems that
think for themselves
I'm Conal Donovan — a Senior Accounting Analyst who got tired of doing the same reconciliation twice. Now I build the systems instead: multi-agent FP&A platforms, survival-modeling ML, route-setting analytics, and the occasional piece of hardware, shipped end to end.
Control room
Everything I've built, running at once
Reconcile
Blackboard · ledger agent
Predict
Predisave · recall risk
Route
BetaLoop · setter feed
Arrive
Transit Board · Pico W
- Python
- Excel + VBA
- RPA
- LLM agents
- Multi-agent orchestration
- Survival analysis
- Forecasting
- TypeScript
- Next.js
- MicroPython
- Raspberry Pi Pico
- E-paper
- Data viz
Selected work
Four problems, four very different systems
Each card is a working miniature of the real thing. Hover it, poke at it (the transit board takes taps), or open it up for the full story.
Multi-agent FP&A platform
Blackboard
A blackboard-architecture system where specialized agents — reconciliation, variance, forecasting, reporting — read and write to a shared state, closing the books without a human in the loop for routine cycles.
Survival modeling for medical recall risk
Predisave
Route-setting analytics platform
BetaLoop
Pico W · e-paper · tap to cycle
Raspberry Pi Pico e-paper display
Transit Board
Playground
This rec doesn't tie. Find out why.
A real (tiny) spreadsheet with a real formula engine. Edit any cell and watch the recalculation ripple through, or let the agent hunt the mistake down the way an accountant would. Then hit New mistake and do it again.
What I do
Four capabilities, one habit of shipping the whole thing
I started in the ledger, not the IDE, and that turns out to be the advantage. I know which reconciliations actually break, which reports nobody reads, and where an agent can safely be trusted with the pen.
Monthly close
Days → hours
- Pull trial balance00:04
- Reconcile bank & subledgers00:21
- Book accruals00:37
- Explain variance vs. forecast01:12
- Draft flux commentary01:58
Senior Accounting Analyst by trade
Engineer by necessity, because spreadsheets weren't cutting it.
Agentic automation
Multi-agent systems that close the books, reconcile ledgers, and flag variance without a human babysitting every step. Built on real accounting workflows, with the audit trail an accountant would actually ask for.
Applied ML & forecasting
Survival models and time-to-event prediction for problems where 'will it happen' matters less than 'when': recall risk, churn, anything with a clock attached.
Data platforms & feedback loops
Analytics that close the loop between what people do and what gets built next. Usage heatmaps, consensus tracking, and dashboards people actually open on Monday.
Embedded & hardware
Raspberry Pi Pico, e-paper displays, MicroPython. Small physical devices that do one job well and don't need an app to do it.
Approach
How I actually work
not just what I build
Automate the boring parts
If a process is tedious and error-prone, that's a signal it should be a system, not a habit. RPA, agentic workflows, and the occasional well-placed Excel macro, whichever is the smallest thing that works.
Design for explainability
A model or agent that can't show its work doesn't get trusted with real decisions. Everything I build leaves a trail a controller could follow, because eventually one will.
Ship the whole stack
From a Pico's firmware to a forecasting model to the dashboard someone opens Monday morning, I'd rather own a thin slice end to end than a thick slice of one layer.
“The best automation is the kind nobody notices. The close just happens, the risk curve is already on the screen, and the transit board already knows you're running late.”
Conal Donovan
Senior Accounting Analyst, systems builder
Got something tedious that should be a system?
Tell me about the spreadsheet you dread every month, or the process everyone agrees is broken but nobody owns.