Manifesto
Funds and insurers are running AI workflows that touch hundreds of decisions a day. Nobody can tell you what any one of them costs to run, why it costs that, or whether the output justified the spend. That gap is not a reporting problem. It is a governance problem.
Founders
Built AI-powered platforms and ML pipelines. ML Engineer at PixAi (generative AI model development) + pipeline dashboard tooling at SIL Global + NLP research at CMU — spanning the full AI engineering stack.
2 years at Caisse des Dépôts (€1.4T Total Assets) modeling multi-billion investment decisions and scenario planning. GTM for a compliant blockchain solution @The Aha Company (Station F).
The team reached the finals of the 2026 HEC/Paris Fintech Hackathon with an AI FinOps SaaS built in 24 hours.
01
100%
Spend attributed per component
02
5 layers
Instrumented per workflow
03
<7 days
From integration to first report
04
0
Vendor lock to a model provider
What we instrument
005 components
01
Model selection
Which model was called, why, and what the equivalent would have cost elsewhere.
02
Prompt construction
Token weight of every assembled prompt, attributed to the workflow that built it.
03
API orchestration
Every chained call, fan-out, retry, and timeout — priced and traced end to end.
04
Retrieval & RAG
Vector store reads, reranks, embedding calls. The line items behind every lookup.
05
History accumulation
Conversational state that quietly compounds. We surface the spend before it bites.
Pipeline · instrumented end to end
Cost is traced
at every node.
00
Input
01
Model selection
↕ Cost · Visibility · Action
02
Prompt construction
↕ Cost · Visibility · Action
03
API / orchestration
↕ Cost · Visibility · Action
04
Retrieval / RAG
↕ Cost · Visibility · Action
05
History accumulation
↕ Cost · Visibility · Action
06
Output
Lens
Cost
What this step charged, in dollars and tokens.
Lens
Visibility
What Firedog exposes here, in real time.
Lens
Action
What you can kill, renegotiate, or reroute.