Melete — find the best answer in the fewest experiments

What it does — one example

You run a coffee shop and want the best espresso. You can change three things — water temperature, grind, and how many grams of coffee. There are thousands of combinations, and testing one means brewing a cup and tasting it. You can't try them all.

Melete works like a brilliant assistant who suggests the next cup to brew:

☕ Melete: “Try 92°, grind 6, 18g.” → you brew it, taste it: 7/10.

☕ Melete: “Now try 93°, grind 5, 19g.” → you taste: 8.5/10.

… a few more …

🎯 After ~20 cups, it found your best recipe — instead of randomly trying 200.

Swap “coffee” for a training run, a chemical reaction, or a price — it's the same: Melete finds the best settings in the fewest expensive tries, and signs a proof of how it got there. You bring the thing you can adjust and a way to score one try; it brings the strategy.

Meet Meli — a tiny story

💛

Once upon a time, a little coffee shop wished for the most delicious espresso in the world.

☕☕☕
☕☕☕

But there were thousands of ways to make it — and every single test meant brewing, and tasting, a whole cup. Trying them all? Impossible.

brew this one →

Then came Meli — who never tries everything. Meli looks, thinks, and the little light glows: “brew this one next.”

☕7 → 8.5 → 9.2

You brew it, you taste it — 7 out of 10. Meli smiles, learns, and picks an even smarter cup. 8.5… 9.2…

⭐📜 verified ✓

In about twenty cups, Meli found the perfect recipe — and sealed a magical proof of how, so the whole world could trust it. The end ✨

▶ Now watch Meli do it for real

How it works — 3 steps

Set the dials

List what you can change and its range — temperature 85–96°, learning-rate 0–0.1, price $1–100.

Score one try

Your real process returns one number: brew → taste, train → accuracy, price → revenue. No dataset needed.

Discover & prove

Melete proposes the next experiment, learns, converges to the best — and signs a verifiable trace of how.

Who it's for & what they get

AI / ML teams

Hyperparameter & system tuning

Tune learning rates, architectures, RAG/serving configs, compiler flags — fewer GPU-hours to the best model, with a provable tuning record.

Pharma · Chemistry · Materials

Formulation & reaction discovery

Find the reagent mix / conditions that maximise yield or potency in far fewer assays — and a tamper-proof discovery trail for patents & audits.

Semiconductor · Manufacturing

Process optimisation

Tune deposition / etch / print parameters against real KPIs on-prem — air-gapped, data never leaves the fab, result still verifiable.

Quant · Product · Growth

Pricing & expensive A/B

Search price points, configurations, and policies where each test is costly — converge faster than grid or manual search.

Every case: fewer expensive experiments to the best answer + a cryptographic proof of how it was found.

See it discover — just watch

Melete tunes knobs. You don't write code or upload data here — pick a scenario and press Watch.

In this browser demo the "score" is faked by a formula so it runs instantly. For your real work the score comes from your real process — see how to use it for your work below.

Scenario

In the browser the score is a simulated model of the process. For real numbers, connect your real process (see below).

Space — the variables (name · type · min · max) Objective — the simulated score (a formula, browser demo only) Budget — experiments allowed

Pick a scenario, then press Watch — the best settings, a movie of how it searched, and a signed proof appear here.

Discovery cinema — watch it search, coloured by strategy

Heat = the score it learned · each dot = one experiment, coloured by the strategy that proposed it · ★ = best.

Convergence

Which strategy the bandit chose

Proof

Click an industry — see Melete work on it

Each card runs the live demo on a realistic, domain-shaped scenario. The browser score is a simulated model of the process — the optimisation is real & reproducible; connect your real assay / benchmark / process for real numbers.

💊 Pharma · biotech

Drug formulation

Variables: pH · temperature · excipient %. Goal: stability / potency. Melete finds the most stable formulation in ~60 assays — instead of hundreds.

▶ Run it now

⚡ AI infrastructure · accelerators

GPU kernel tuning

Variables: tile size · unroll · occupancy. Goal: throughput (GFLOP/s). Find the fastest config in ~50 benchmark runs.

▶ Run it now

🔬 Semiconductor · fab

Plasma-etch process

Variables: power · pressure · time. Goal: wafer yield %. Tune the recipe to maximum yield — air-gapped, on-prem.

▶ Run it now

🧠 The AI world itself

LLM serving config

Variables: batch size · KV-cache · quantization. Goal: tokens/sec at a quality bar. Melete optimises AI infrastructure too — and can tune prompts, agents & routing the same way.

▶ Run it now

☕ Everyday

Best espresso recipe

Variables: temp · grind · dose. Goal: taste. The friendliest way to watch the idea click.

▶ Run it now

Want this for your team — or to acquire the code?

Melete is a clean, dependency-free codebase with a live demo, full tests, and a signed-provenance moat. Open to licensing or an IP acquisition.

📩 Contact about Melete Read the pitch

Proven, not claimed

No single optimiser wins on every landscape. A bandit spends each experiment on whichever strategy is winning on your problem — one engine, no per-problem tuning.

landscape	Melete	single Bayesian	random
smooth	1.000	0.999	0.838
rugged (many traps)	best 🏆 beats every single algorithm	far behind	far behind
high-dimensional	0.996	0.987	0.555

≈ 26 experiments vs ~95 for random to reach 99% of a hidden optimum (3.7×). Reproduce with melete bench --robust.

Use it for your work — answer 3 questions

No dataset, no formula. Just answer these about your process:

What can you adjust?

List the knobs + their real limits (your machine's range). → that's the SPACE.

After one try, what number tells you how good it was?

You measure it — taste a score, read accuracy, read revenue. You don't calculate it. → that's the SCORE.

How many tries can you afford?

Brews, training runs, assays you'll pay for. → that's the BUDGET.

☕ A coffee shop

Knobs: temp 85–96° · grind 1–10 · dose 14–22g
Score: a barista tastes each shot, 0–10
Budget: 30 shots → Melete finds the recipe in ~20.

🤖 An ML team

Knobs: learning-rate 0–0.1 · depth 1–12
Score: the training script prints accuracy
Budget: 40 runs → fewer GPU-hours to the best model.

Then run it one of two ways:

A) Connect your process — Melete runs it for you and reads the number (this is the real product, like installing a tool):

melete tune --cmd "python train.py --lr {lr} --depth {depth}" \
            --space '[{"name":"lr","type":"real","min":0,"max":0.1},{"name":"depth","type":"int","min":1,"max":12}]'

B) From an agent or pipeline — call the HTTP API or the library; your code returns the score each step:

POST https://melete.mneme-ai.space/discover   ·   npm i melete-ai   ·   discoverSigned({ space, oracle })

This website = a sandbox to try it. Real work = connect your real process (A or B). 🔒 Air-gapped: zero dependencies + local signing ⇒ runs fully offline, result still verifiable.