MonceApp is a GenAI platform with 14 live models and 3 lazy constructors (Charles, Moncey, Json),
serving 100 concurrent users on a $15/month EC2. The SDK (pip install monceai) requires zero API key.
Arithmetic and glass decomposition bypass LLMs entirely via regex → boolean circuits → SAT solvers.
The system integrates 10 computation services, uses Snake SAT for routing, and binary multiplier/divider
circuits for zero-token exact arithmetic.
Pure function: str prompt → str enhanced_prompt. Injects factory-specific glass knowledge,
IGU rules, coating normalization, and operator context. Frameworks registered via POST /v1/framework.
pip install git+https://github.com/Monce-AI/monceai-sdk.git
| Constructor | Type | What |
|---|---|---|
Charles("6x7") | str | Math, science, general — lazy parallel, resolves on read |
Moncey("44.2 feuillete") | str | Glass sales agent — snake + classifiers + Haiku |
Json("list 5 primes") | dict | Structured output — dict subclass, json.dumps(indent=2) |
Json("..." + Moncey("...")) | dict | Chains: Moncey resolves, Json structures |
| Model | Engine | Cost/msg | Latency |
|---|---|---|---|
| charles | 4x parallel (haiku mem + csv + cnf + sudoku) → Sonnet | ~$0.01 | 8-15s |
| charles-science | Haiku features → Snake router → 7 services → Sonnet | ~$0.01 | 15-60s |
| charles-auma | Haiku encode → AUMA {0,1}^n → Haiku (= or ≈) | ~$0.003 | 3-8s |
| concise | charles → Haiku TL;DR (420 tokens) | ~$0.01 | 10-20s |
| cc | charles ∥ concise → Haiku synthesis | ~$0.02 | 12-25s |
| charles-json | Memory (5s cap) → Sonnet strict JSON, VLM | ~$0.01 | 5-15s |
| charles-architect | Memory → Sonnet ASCII diagrams | ~$0.01 | 5-15s |
| Sonnet 4.6 | Bedrock Converse + tools | ~$0.03 | 1-3s |
| Sonnet 4 | Bedrock Converse + tools | ~$0.03 | 2-4s |
| Haiku 4.5 | Bedrock Converse + tools | ~$0.003 | 1-2s |
| Nova Pro | Bedrock Converse (context only) | ~$0.008 | 0.8s |
| Nova Lite | Bedrock Converse (context only) | ~$0.001 | 0.7s |
| Nova Micro | Bedrock Converse (context only) | ~$0.0005 | 0.6s |
charles-auma has three execution paths, fastest wins:
A×B encoded as carry-chain adder CNF: fix A,B bits as unit clauses, Kissat propagates to product bits. Division is the inverse: fix A,B, solve Q×B + R = A for Q and R.
| Problem | Answer | Vars | Time | Tokens |
|---|---|---|---|---|
| 6×7 | 42 | 27 | 97ms | 0 |
| 123×456 | 56,088 | 187 | 55ms | 0 |
| 65535×65535 | 4,294,836,225 | 768 | 58ms | 0 |
| 42÷7 | 6 r 0 | 45 | 82ms | 0 |
| 1000÷7 | 142 r 6 | 95 | 49ms | 0 |
| 998001÷999 | 999 r 0 | 353 | 48ms | 0 |
| Problem | Answer | Bits | AUMA | Evals |
|---|---|---|---|---|
| 6×7 | = 42 | 6 | bits [1,0,1,0,1,0] | 1,296 |
| 999² | = 998,001 | 20 | exact | 160,000 |
| 40-bit | = exact | 40 | exact | 2,560,000 |
| Polynomial | AUMA found | Exact | Time |
|---|---|---|---|
| x³ − 6x + 4 = 0 | x ≈ 0.733 | 0.732 | 0.6ms |
| x² − 7x + 12 = 0 | x ≈ 2.98 | 3 | 0.4ms |
| z² + 1 = 0 | z ≈ −0.005 + 0.992i | ±i | 3.3ms |
| z&sup5; = 1 | z ≈ 1.003 | 1 (principal) | 3.4ms |
A Snake classifier (42 training samples, 15 layers, trained in 0.01s) routes prompts to the right service. Haiku extracts 16 boolean features from the prompt, Snake classifies instantly (0ms, 100% confidence on all tested cases), then Haiku extracts parameters for the selected service.
| Service | URL | What | Latency |
|---|---|---|---|
| AUMA | auma.aws.monce.ai | Universal maximization, polynomial budget | <100ms |
| Sudoku | sudoku.aws.monce.ai | Polynomial BCP solver | 0.1-500ms |
| NPDollars | npdollars.aws.monce.ai | Swarm SAT solver | 5ms-60s |
| SnakeBatch | snakebatch.aws.monce.ai | Distributed Snake training | 0.6-20s |
| Chess | chess.aws.monce.ai | AUMA chess engine vs Stockfish | 50ms-6s |
| Snake | snake.aws.monce.ai | Glass article SAT matching | 0.15ms |
| KPI | data.aws.monce.ai | Live extraction accuracy/volume | 2-5s |
Deterministic-first, LLM-second. Like NP calc for arithmetic, snake/comprendre decomposes glass without LLM.
from monceai import Moncey, Json
Json("Extract order: " + Moncey("44.2 Silence/16 alu/4 rFloat JPP"))
# → {"articles": [{"name": "Feuillete 44.2", "ref": "#60442"}, ...]}
| Step | Service | What | Latency |
|---|---|---|---|
| 1 | snake.aws.monce.ai/comprendre | Glass decomposition: verre, intercalaire, gas | 224ms |
| 2 | moncesuite.aws.monce.ai/comprendre | 10 classifiers (only if snake quality < 75%) | 206ms |
| 3 | Haiku | French sales response with context | ~3s |
Snake SAT Memory: glass terms matched against local Snake classifiers (~50ms). NP-Complete Arithmetic: all formulas intercepted and computed exactly via AST-safe eval. charles-auma skips NP calc — AUMA is the calculator, finding answers from bits.
Snake SAT classifiers are built on the Dana Theorem (2024): any indicator function over a finite discrete domain can be encoded as a SAT instance in polynomial time. Snake is also the router — classifying prompt features into service selections at zero token cost.
charles-auma implements the AUMA thesis directly: any function f: {0,1}n → R can be maximized by Fourier probing + greedy search in O(na) evaluations. Integer arithmetic is a special case where the answer's binary representation IS the {0,1}n point that maximizes −(decoded − target)².