Roadmap

How we build, in the open.

Falcron ships production systems, not slideware. This is where we are, what we're building now, and where we're headed — framed as direction, not dated promises.

Shipped — in production Building — live work Next — committed direction
Shipped — live today
Phase 01

Foundations in production

The data layer and interfaces our clients run on right now.

  • Multi-source data lakehouse — automated bronze → silver → gold pipelines turning raw operational exports into clean, queryable gold tables. 20M+ rows live across deployments.
  • Live operational dashboards — decision-first, reading directly from gold, updated on every pipeline run.
  • Text-to-SQL / natural-language data access — ask in plain English, get a real answer over the lakehouse.
  • Autonomous ingestion & RPA bots — unattended, scheduled, with zero-error production runs.
  • Private / local-first deployment — pipelines and agents that run inside the client's own environment when data can't leave the building.
Outcome

Clients move from manual spreadsheet exports to a single, trusted, queryable source of truth — live, automated, and private.

Building — in active development
Phase 02

Production-grade AI ordering agents

Autonomous agents that decide and place stock orders for pharmacies — trained on real operational history and gated by independent verification before a single order is committed.

  • Trained on ~20M rows of operational history — dispensing, purchasing, stock movement and expiry data across multiple pharmacy sites — so forecasts reflect how each branch actually buys, not a generic model.
  • Per-pharmacy demand & stock-optimisation models — patterns differ branch to branch — project what each line needs over the lead-time window, accounting for seasonality, concession and tariff shifts, and recent dispensing trend.
  • Multi-agent juror verification — each proposed order is reviewed by several specialised agents that independently check demand, wholesaler choice, forward-buying logic and margin. An order only proceeds if the panel agrees; no single model can place an order alone.
  • Hard guardrails enforced before commit — budget ceilings, expiry / shelf-life limits, shortage and over-order protection, and supplier availability. Any breach blocks the order regardless of agent confidence.
  • Human-in-the-loop sign-off with a full audit trail — every order surfaces its forecast, the agents' reasoning and the juror verdict for one-click approval, and every decision is logged immutably.
  • Agent observability built in — per-order decision tracing, token and cost accounting, and a record of which agent proposed, which dissented and why.
Outcome

Ordering decisions that are explainable, reviewable and safe by construction — with a human approving and a complete trail behind every order.

~20M rows
operational history the agents train on
Multiple jurors
independent agents verify every order
£0
placed without guardrails + human sign-off
The juror approach

Because every order moves real money and affects patient supply, we never let one model decide alone. Each proposed order is reviewed by several specialised agents — a jury — that independently scrutinise the forecast, the wholesaler choice and the margin impact. The order only goes forward if the panel agrees and every hard guardrail passes; any dissent or breach sends it back or escalates it to a human. Adversarial verification by design: nothing executes unless it survives independent review.

Juror verification · liveevaluating…
Proposed order
#4471 · 142 lines
£3,820 · wholesaler routed
Demand
forecast within tolerance
Wholesaler
best price · in stock
Margin
above floor
Expiry
within shelf-life
Budget
under ceiling
Human
sign-off

No single model places an order. The panel must agree and every guardrail must pass.

Next — committed direction
Phase 03

Scaling ordering agents into a platform

Take the verified ordering agents from a single-group deployment to a self-serve platform any pharmacy group can run on its own data.

  • Multi-pharmacy orchestration — one control plane running independent ordering agents across many sites and groups, each isolated to its own data while sharing the same verified engine.
  • Self-serve onboarding — a new group connects its data and wholesaler accounts, the agents calibrate on its history, and the system reaches supervised ordering in days, not a bespoke build.
  • A library of evaluated, version-pinned agents — ordering strategies, juror / verification roles and wholesaler connectors, each scored on back-tested data so operators can see how a component performs before enabling it.
  • Expanded wholesaler connectors covering ordering, pricing and availability, so agents can compare and route across suppliers automatically.
  • Privacy-preserving cross-pharmacy benchmarking — anonymised, aggregated signals let each group see how its buying, margin and shortage exposure compare to peers, without exposing any other pharmacy's raw data.
Outcome

A productised ordering-agent platform new pharmacy groups can adopt safely — with benchmarking that gets sharper as more sites join.

We're a small, senior team. We list direction, not delivery dates — and we only move work into Shipped once it's running in production for a real client.

Ready to ship

Let's get your production AI live.

Bring us your operational data. We'll build the system that acts on it — deployed in your environment, measured in production.