Stacks/Fintech / internal platform
Bkper's AI gateway, app-building agent and Gemini Flash parsers (X)
Bkper founder Mael Caldas describes 18 months of agents in production: an in-house AI gateway, CLI harness, and sandboxed managed agent that users use to build and deploy apps. For bookkeeping operations, LLM agent loops didn't fit (accounting has to be 100% right), so what works is narrow: Gemini Flash Lite extracts stable parts of a transaction description to find account candidates, and Gemini Flash structured outputs classify and parse uploaded documents, with prompts that self-adjust from user corrections. They now route judgment-model calls (TypeSafe Jev) through the gateway behind their own open spec.
- Use case
- An AI-first bookkeeping platform. A managed agent lets users build and deploy their own apps in an ephemeral Linux sandbox, while bookkeeping operations use narrow AI steps: account discovery for transactions and document classification/extraction for invoices, receipts, and statements.
- Category
- Fintech / internal platform
- Team size
- not stated (team in Brazil handles operations and clients)
- Orchestrator / framework
- Their own AI gateway, CLI harness, agent sandbox, and app-building platform. Operations are driven by events and structured outputs rather than agent loops; judgment-model calls sit behind a standard gateway interface with adapters so the underlying model can be swapped without changing client code.
- Memory
- The managed agent's context is managed by the user, bounded by the platform's security and distribution model. Learned prompt adjustments from user corrections are saved in properties and reused for the next parsing.
- Evals
- not stated
- Observability
- not stated
- Cost (stated in writeup)
- Not stated
- Cost per output (stated)
- Not stated in the writeup
- WhatWeRun estimate
WhatWeRun estimate
Not enough info to estimate
Models are named but the article gives no token counts, volumes, or per-task workload, and judgment-model pricing isn't stated.
Reviewed Nov 14, 2026
- Models
- Observability tools
- not stated
- Other tools
- Ephemeral Linux sandbox per managed agentIndexed base of accounts for candidate lookupSpecific OCR parsers for one large client
- Source
- https://x.com/maelcaldas/status/2121042288315134246
Mael Caldas (@maelcaldas), founder of Bkper · Published Nov 12, 2026 · Last verified Nov 14, 2026
Related
More like this
Stripe Radar's Agentic MLE
AI agents that run ML experiments on Shield, Radar's transaction-fraud model: investigating missed fraud, implementing candidate features, training and evaluating models against a baseline
zenml.ioBlock KGoose (Money Bot, Manager Bot, and G2)
A shared AI agent platform at Block that powers Money Bot (a Cash App financial assistant for peer-to-peer payments), Manager Bot (a Square merchant assistant), and G2 (an internal productivity platform)