Case studyFounder and engineer2026-02 – present

Radicas

A FinOps platform that measures what an agentic AI feature costs and what it returns, by reconciling real-time OpenTelemetry spans against the invoice that arrives days later.

As of October 2026.

Customer names and revenue figures is not public — commercially confidential.

Ventobacked by Vento, Exor's early-stage VC armSelf-reported; no independent source.
5 moproblem to first usersSelf-reported; no independent source.
7agent frameworks on one data model, 6 experimentalSelf-reported; no independent source.

How it works

Diagram generated from the knowledge base
How Radicas reconciles cost OpenTelemetry gen_ai spans are priced in real time from a price book, giving an estimated cost per call. The provider's invoice arrives days later as FOCUS 1.3 data, and each span is reconciled against its invoice line items, so the real-time estimate is corrected by the authoritative figure. The result is cost per feature and per tenant. gen_ai spansfrom OpenTelemetry Real-time estimatepriced from a price book Provider invoiceFOCUS 1.3, days later Reconciliationspan by span against line items Cost per featureand per tenant

Context

Teams shipping agentic AI features get a cloud bill and a token count, neither of which answers the question the business asks: what does this feature cost per use, and is it worth it. Provider consoles aggregate by account or by key, not by the feature or the tenant that caused the spend. Radicas's public site now frames it as AI cost management and governance for work done by people and agents, measured per unit of work.

Problem

Two sources of truth disagree by design. Instrumentation produces a cost estimate at the moment a call happens, computed from a price book. The provider's invoice is authoritative but lands a day or three later, and it is grouped by nothing the product cares about. Using only the estimate gives numbers nobody will sign off on; waiting for the invoice gives numbers that arrive too late to act on.

Role

Founder and engineer. Problem discovery, architecture, implementation and go-to-market, inside the five-month Reef by Vento programme. That includes the Python SDK and a Claude Code setup plugin.

Decisions

The reasoning behind these — the alternatives weighed and why they lost — is withheld while the product is live; see the withheld note in this page's metadata.

01 · ClickHouse, LLM observability and tracingReconcile spans against invoice line items, and make that the productClickHouse, LLM observability and tracing

Ingest OpenTelemetry gen_ai spans, price them in real time from a price book, then reconcile them span by span against FOCUS 1.3 invoice data when it arrives. The join is the core of the product, not a reporting feature on top of it.

Source code is not public — proprietary.

Design rationale — the alternatives weighed and why they were rejected is not public — commercially sensitive while the product is live.

02 · TerraformEU-native hostingTerraform

Infrastructure is defined in Helm and Terraform, with data residency in the EU.

Source code is not public — proprietary.

Design rationale — the alternatives weighed and why they were rejected is not public — commercially sensitive while the product is live.

03 · Multi-tenant SaaS architectureOne artifact for SaaS, self-host and BYOCMulti-tenant SaaS architecture

The same build ships as managed SaaS, as a self-hosted install and into a customer's own cloud, with one upgrade path across the three.

Source code is not public — proprietary.

Design rationale — the alternatives weighed and why they were rejected is not public — commercially sensitive while the product is live.

04 · LLM observability and tracingOne data model across agent frameworksLLM observability and tracing

A card per framework maps its spans onto the OpenTelemetry gen_ai span tree and generates three layers: collector transform rules, an SDK adapter and a query-time fallback. Each recognised span carries its source framework. Seven agent frameworks and four model-provider sources are covered; Google ADK is verified end to end, the rest are experimental.

Source code is not public — proprietary.

Design rationale — the alternatives weighed and why they were rejected is not public — commercially sensitive while the product is live.

Details

Organisation
Reef by Vento
Skills
Agent-assisted deliveryFinOps for AILLM observability and tracingPython and FastAPIGoClickHouseTerraformMulti-tenant SaaS architectureDeveloper toolingMCP servers

Results

Assessed as of the date in the front matter; the product is in active development.

  • A local end-to-end lab runs the reconciliation, joining per-span estimates to FOCUS-shaped line items on the span id. Those line items are seeded from the lab's own telemetry, not a provider invoice.
  • A Claude Code setup plugin is built and in pre-release, with a CLI and a GitHub Action alongside it. Its setup skill finds a repository's Claude Code agents in GitHub Actions, stores the ingest key as a CI secret without it entering the conversation, opens a pull request that wires them up, and verifies the first run; the user approves each step.
  • Around thirty ADRs record the reasoning behind the architecture.
  • Around thirty ADRs are written down behind the product, each with its reasoning and what would change it.

    you can check this: Francesco Fiore on GitHub

  • Went from problem to first users inside the five-month Reef by Vento programme.

    self-reported; no independent source

  • Built and backed inside Reef, the venture-building programme of Vento, Exor's early-stage VC arm.

    self-reported; no independent source

  • The Python SDK maps seven agent frameworks onto one OpenTelemetry gen_ai data model; Google ADK is verified end to end, the other six are experimental.

    self-reported; no independent source

RadicasRadicas codebase