Autonomous Optimisation
as a Service.
Your operations have optimal parameters. You don't know them yet. We build a simulator from your historical data, deploy an autonomous agent, and run experiments until the answer is found.
Build. Simulate.
Optimise.
We apply the autoresearch pattern, developed by Andrej Karpathy, to business operations. Define a precise metric. Build a fast simulator from your historical data. Deploy an autonomous AI agent to run experiments: adjust parameters, measure results, keep what works, revert what doesn't.
Automated experimentation at machine speed. Near-optimal parameters found without guesswork. The same pattern used in frontier AI research, applied directly to your operations.
Real experiments.
Real results.
The same autonomous optimisation pattern validated across two domains — inventory and pricing — at escalating levels of complexity. Each version is harder. Each version proves the pattern holds.
5 products · Fixed lead times · Simple stockout and holding cost formula. The proof-of-concept that started the pattern.
12 products · Per-product lead times (1–5 days) · Quantity discount tiers · 36-parameter search space. The agent discovered sharp cliff boundaries where a 1-unit change caused an 8× jump in missed sales.
12 products + perishable goods · FIFO batch tracking · Variable shelf lives (3–20 days) · Bayesian GP surrogate · Gradient-guided search. Same improvement, 77% fewer experiments, on a harder problem.
Time-phased per-product strategies · Per-product Bayesian GP (3D) · 15,000 experiments in 15 minutes. The AI stopped running experiments and started writing automated optimisation pipelines. It also identified the structural cost floor.
The V2 to V4 progression tells the real story. V2 proved the pattern on realistic complexity. V3 showed that harder problems don't require proportionally more effort when you invest in smarter optimisation. V4 pushed to the structural cost floor: 34% of total cost is unfixable by parameter tuning alone. The AI identified which products have shelf_life approximately equal to lead_time and flagged them for operational intervention, not further experimentation. That is the difference between a tool that optimises and one that advises.
8 products · Price elasticity + demand cliffs · Volume discount tiers. The agent pushed every product to its exact cliff threshold — the highest price before demand collapses.
Cross-price elasticity across substitute groups · 3 product families. When premium_widget's price rose, demand shifted to standard_widget. The agent learned to optimise group-level dynamics.
3 customer segments · Seasonal demand multipliers · Competitor price reactions · 32-parameter search space. Monthly differentiation was tested exhaustively and reverted every time — cliff thresholds dominate.
Monthly pricing · 3 product bundles · 99 parameters · £578k → £1,027k. The agent discovered that pricing budget_widget at cost (a £71k individual loss) generates £482k in bundle revenue. The loss leader is portfolio optimisation.
V4's 77.5% improvement didn't come from better optimisation of the same problem — it came from bundles creating a new profit mechanism. The agent discovered that deliberately losing money on one product (budget_widget at cost) drives enough bundle volume to generate £482,000 in bundle revenue. 118 experiments outperformed V3's 23,522 because the agent had learned from V1–V3 that cliffs are the dominant constraint, and jumped straight to the loss-leader discovery once bundles were available.
Five domains.
One repeatable pattern.
The autoresearch pattern is domain-agnostic. Anywhere you have a measurable objective, historical data, and controllable parameters, the same approach applies. Validated on inventory. Deploying across four more.
Reorder points, safety stock, and order quantities optimised against your actual demand patterns and cost structure. Full perishable goods support with FIFO batch tracking, shelf-life-aware ordering, and waste-vs-stockout trade-off analysis.
Price elasticity modelling and optimal price point discovery across your product catalogue. We fit demand curves from historical sales data, then let the agent find prices that maximise revenue or margin.
Diminishing returns modelling per channel with optimised budget splits. We replace platform-reported ROAS with independently verified allocation across Google, Meta, and beyond.
Shift staffing matched to demand patterns, minimising labour cost while meeting service level targets. Built for restaurants, retail, call centres, and any business with variable-demand shift work.
Route sequence optimisation for delivery companies, field service, and last-mile operators. Minimise total distance, time, and fuel cost while respecting vehicle capacity and delivery time windows.
Done for you.
Delivered in a week.
A consulting engagement. You share your data. We scope the problem, build the simulator from your historical figures, run the autonomous optimisation loop, and deliver actionable parameter recommendations with a full experiment log. Ready to implement in your existing workflow. No new software required.
Start with one problem.
Share your data. Define the metric. We scope the engagement in one call and tell you exactly whether the pattern applies and by how much.