TECHNOLOGY INTELLIGENCE PLATFORM
RESTRICTED — AUTHORISED PERSONNEL ONLY
ALPHANUMERIC
TECHNOLOGY INTELLIGENCE PLATFORM
CONFIDENTIAL — AUTHORISED VIEWING ONLY
REPORT SUMMARY MERQUAN TECHNOLOGY INTELLIGENCE  |  2026-05-06  |  CONFIDENTIAL
CategoryCountDetails
NVIDIA Libraries15+CUDA 12.8 full suite + cuQuantum + NCCL + Triton
LLM Providers4Anthropic · DeepSeek · NVIDIA NIM · Novita (proxy)
LLM Models Available13+Claude S4.6/H4.5/O4.7 · Nemotron 49B (self-hosted)/70B · Llama 405B/3.3-70B · Maverick 17B · DeepSeek V3/V4 · FinBERT
GPU Frameworks8JAX/XLA · PennyLane · cuPy · PyTorch · WebGPU
ML Libraries8scikit-learn · SciPy · NumPy · Numba · SymPy · QuantEcon · Transformers · CCXT
Market Data Sources9Rithmic · Databento · Finnhub · OANDA · FRED · CME · FIGI · CCXT · TV Scanner
Storage Layers3Redis 7.4 · QuestDB 9.3.4 · SQLite (quantumx.db / correlations.db / atallum.db)
Active Microservices10+FastAPI · Django · TankFarm · Celery · Rithmic · ExecBridge · QuantumTunnel · nginx
GPU Servers7SaladCloud JAX (Tier 1 ML) · SaladCloud Nemotron 49B (deploying) · H200 (on-demand) · Main server CPU
Observability Tools4W&B Weave · wandb · Sentry · OpenTelemetry
Cloud Providers3DigitalOcean · Google Cloud · SaladCloud
Total Python Packages172+172+ packages — plus esig, roughpy, pennylane-lightning, lineax, jaxtyping (added May 2026)
REPORT HARDWARE RANKED BY GPU POWER  |  CONFIDENTIAL
RankGPUVRAMArchitectureMemory BWLocation / UseTier
#1 ◆ FLAGSHIPNVIDIA H200 SXM141 GB HBM3eHopper GH100 | 16,896 CUDA cores | 640 Tensor cores3.35 TB/sDedicated — MERQUAN COMMAND ssh h200 | Tailscale 100.117.88.107 HPC SDK 26.3 | cuQuantum | RAPIDS | cuOptDEDICATED
#2NVIDIA A100 SXM × 280 GB HBM2e × 2 = 160 GB totalAmpere GA100 · 6,912 CUDA · 432 Tensor · 108 SMs2 TB/s × 2MERQUAN-R1 Training / NeMo Blueprint / data-flywheelTRAINING
#3NVIDIA H100 SXM5 (NIM cloud)80 GB HBM2eHopper GH100 — 16,896 CUDA cores, 528 Tensor cores3.35 TB/sNVIDIA NIM Cloud integrate.api.nvidia.com/v1 Per-token via Enterprise AccountNIM CLOUD
#4 FALLBACKNVIDIA L4048 GB GDDR6Ada Lovelace AD102 | 18,176 CUDA | 568 Tensor864 GB/sPrevious dedicated GPU server Fallback / overflow computeFALLBACK
#5NVIDIA RTX 509032 GB GDDR7Blackwell GB202 | 21,760 CUDA | 680 Tensor1,792 GB/sSaladCloud pool — best nodeDEDICATED
#6NVIDIA RTX 409032 GB GDDR6XAda Lovelace AD102 | 16,384 CUDA | 512 Tensor1,008 GB/sSaladCloud poolDEDICATED
#7NVIDIA RTX 3090 Ti32 GB GDDR6XAmpere GA102 | 10,752 CUDA | 336 Tensor936 GB/sSaladCloud poolDEDICATED
#8NVIDIA RTX A500032 GB GDDR6Ampere GA102 | 8,192 CUDA | 256 Tensor768 GB/sSaladCloud poolDEDICATED
#9NVIDIA RTX 309032 GB GDDR6XAmpere GA102 | 10,496 CUDA | 328 Tensor936 GB/sSaladCloud poolDEDICATED
REPORT INFRASTRUCTURE & SERVICES CONFIDENTIAL
▌ CLOUD CONTAINERS
ContainerGPU AllocationCPU / RAM / DiskModel / EngineEndpointPoolStatus
merquanjaxsalad (JAX Tier 1 ML) Best available node from pool RTX 5090 → 4090 → 3090Ti → A5000 → 3090 32 vCPU | 60 GB RAM | 250 GB SSD Priority: High | Replicas: 1 JAX v2 — 9 engines: Black-76 · HAR-RV/GARCH · VaR Path Signatures · Rough Vol Mamba SSM · Neural SDE Deep Hedging · Quantum VQC durian-alfalfa-sz0nun7i8h0cf8ud .salad.cloud:8888 HighACTIVE
NIM Nemotron 49B (LLM Inference) RTX 5090 (32 GB) — required for 49B int4 Fails over if unavailable 32 vCPU | 60 GB RAM | 250 GB SSD Priority: High | Replicas: 1 nvidia/llama-3.3-nemotron-super-49b-v1.5 Self-hosted — replaces cloud NIM API 195.181.163.241:20608 (SSH) Gateway URL: TBD on deploy HighACTIVE
▌ MAIN SERVERS & SERVICES
ServiceStack / VersionCPU / RAMRoleAddress / PortUptimeStatus
MERQUAN Main ServerDigitalOcean VPS — Ubuntu 248 vCPU | 16 GBAll backend services + nginx + SSL161.35.43.80 / merquan.com24/7ACTIVE
FastAPI — merquan_fastapiuvicorn + uvloop | Python 3.121 workerNAWA 14-engine, all trading APIs, SAFA127.0.0.1:500124/7ACTIVE
Django — merquanglobalgunicorn | 8w × 4t | Python 3.118 vCPU shareAPI router, Sanctum admin, CLAW, Celery127.0.0.1:800024/7ACTIVE
Tank Farm — merqintelFastAPI | Python 3.11Candle normaliser, 6 data tanks, QuestDB proxy127.0.0.1:500324/7ACTIVE
RedisRedis 7.4.0In-memoryTick cache, Rithmic DOM store, Celery broker127.0.0.1:637924/7ACTIVE
QuestDBQuestDB 9.3.4Disk: 1M+ barsTime-series OHLCV — 16yr GC + Rithmic 327 contracts127.0.0.1:900024/7ACTIVE
Rithmic R|ProtocolProtobuf/gRPC WSS50+ proto defsLive CME futures ticks, DOM, BBO depthmerquan-rithmic.service24/7ACTIVE
Exec BridgeFastAPI | PythonOANDA demo order execution127.0.0.1:500424/7ACTIVE
Celery + BeatRedis broker | 4 workersBackground tasks, rvol refresh, QuestDB backfillmerquan-celery.service24/7ACTIVE
DO Agent ServerDigitalOcean — London4 vCPU | 8 GBOpenClaw agent, @yasin_scholar_bot, skill sync 30min161.35.43.8024/7ACTIVE
nginxReverse proxy | SSLAll services + static + SSL terminationmerquan.com / merqintel.com24/7ACTIVE
Databentodatabento 0.75.010M 1min bars16yr historical GC — Parquet → H200 pipelineapi.databento.comOn-demandACTIVE
REPORT NVIDIA & CUDA CONFIDENTIAL
LibraryVersionUsed ForStatus
CUDA Toolkit12.8Low-level GPU compute foundationActive
cuDNN9.10.2.21Neural network layer accelerationActive
cuBLAS12.8GPU linear algebra (matrix multiply)Active
cuFFT12.8Fast Fourier transforms on GPUActive
cuSolver12.8GPU-side linear solvers / eigenvalueActive
cuSPARSE12.8Sparse matrix operationsActive
cuRand12.8GPU random number generation (Monte Carlo)Active
cuFile12.8GPU Direct Storage I/OActive
nvJitLink12.8JIT kernel linkingActive
nvTX12.8GPU profiling markersActive
cuQuantum26.01.0Quantum circuit simulation + HMM regime detectionActive
NCCL2.27.5Multi-GPU communicationActive
Triton (NVIDIA)3.6.0GPU kernel compilation + PyTorch JIT fusionActive
NVIDIA NIM Cloud APIHosted LLM inference (Nemotron, Llama, DeepSeek)Active
NIM Self-Hosted (new)llama-3.3-nemotron-super-49b-v1.5 on SaladCloudDeploying
REPORT NIM SERVERS CONFIDENTIAL
TypeOfficial Model NameGPU SpecContainer / ServerPurposeStatus
Cloud NIM API (current)meta/llama-4-maverick-17b-128e-instructNVIDIA Cloud (hosted)integrate.api.nvidia.com/v1Trading bias, CLAW fallbackACTIVE
Cloud NIM APInvidia/llama-3.1-nemotron-70b-instructNVIDIA Cloud (hosted)integrate.api.nvidia.com/v1CLAW institutional inferenceACTIVE
Cloud NIM APImeta/llama-3.3-70b-instructNVIDIA Cloud (hosted)integrate.api.nvidia.com/v1General inference via CLAWACTIVE
Cloud NIM APImeta/llama-3.1-405b-instructNVIDIA Cloud (hosted)integrate.api.nvidia.com/v1Heavy reasoning tasksACTIVE
Self-Hosted NIM (NEW)nvidia/llama-3.3-nemotron-super-49b-v1.5SaladCloud GPU Pool 32 GB VRAM | 32 vCPU | 60 GB RAM195.181.163.241:20608 SaladCloud containerSelf-hosted 49B reasoning — replaces all cloud NIM callsACTIVE
REPORT TIER 1 ML ENGINES NOBODY CAN BEAT  |  CONFIDENTIAL
EngineAcademic SourceWhat It DoesJAX StackEndpointFileStatus
Black-76 Batch PricerBlack (1976) — Futures OptionsFull strike chain pricing + exact Greeks (Delta/Vega/Theta) in one GPU calljax.vmap + jax.grad + jax.jit/api/jax/optionsengines/options.pyLIVE ✓
HAR-RV + GARCHCorsi (2009) + Bollerslev (1986)Heterogeneous autoregressive realised volatility + GARCH(1,1) batch across instrumentsjax.lax.scan + jax.vmap/api/jax/har_rvengines/har_rv.pyLIVE ✓
Monte Carlo VaRBasel II/III10K bootstrap simulations, JIT-compiled, configurable confidence intervaljax.jit + jax.vmap/api/jax/varengines/har_rv.pyLIVE ✓
Path SignaturesLyons (2014) — Oxford Math InstProvably optimal feature extraction. Universal approximation theorem for path-dependent functionalsesig 1.0.0 + numpy/api/jax/v2/signaturesengines/path_signatures.pyLIVE ✓
Rough Volatility (RFSV)Gatheral/Jaisson/Rosenbaum (2018)H≈0.1 Hurst exponent — best vol model known to science. rBergomi simulation + implied vol inversionCustom JAX + jnp.linalg.lstsq/api/jax/v2/rough_volengines/rough_vol.pyLIVE ✓
Mamba SSMGu & Dao (2023) — Mamba PaperO(n) selective state space model. Beats Transformers on long financial sequences. 5-regime detection + directionEquinox 0.13.8 + jax.lax.scan/api/jax/v2/regimeengines/mamba.pyLIVE ✓
Neural SDE (Latent)Kidger et al. (2021) — ICLRDrift + diffusion as neural networks. Full price path distribution (not point estimate). ELBO trainingDiffrax 0.7.2 + lineax + Equinox/api/jax/v2/nsde_pathsengines/neural_sde.pyLIVE ✓
Deep HedgingBuehler et al. (2019) — J.P.MorganEnd-to-end RL hedging strategy. CVaR loss. Handles transaction costs. Replaces Black-Scholes deltaEquinox GRU + jax.lax.scan/api/jax/v2/deep_hedgeengines/deep_hedging.pyLIVE ✓
Quantum-Classical VQCFarhi et al. (2014) QAOA + PennyLaneVQC portfolio optimisation. QAOA ansatz, parameter-shift gradients. Explores Hilbert space for weight optimaPennyLane 0.44.1 + Optax Adam/api/jax/v2/quantum_portfolioengines/quantum_hybrid.pyLIVE ✓
REPORT ML & STATISTICAL CONFIDENTIAL
LibraryVersionAlgorithms / Features UsedUsed In
scikit-learn1.8.0RandomForest, KNN, LogisticRegression, GaussianNB, SVM, PCA, StandardScalermerquan_algo.py, merquan_arb_engine.py, merquan_ml_engine.py
SciPy1.17.1stats (norm, entropy, linregress), optimize (minimize, linprog), signal (savgol_filter), cluster.hierarchyPhysics engines, portfolio optimisation
NumPy2.4.3Core arrays, ORJson numpy serialisation (OPT_SERIALIZE_NUMPY)Throughout — all engines
Numba0.65.0@njit JIT compilation for tight-loop indicators (moving averages)merquan_engines_fast.py
SymPy1.14.0Symbolic mathematics in physics enginesPhysics engine modules
QuantEcon0.11.2Economic models (Markov, LQ control, etc.)Engine research modules
HuggingFace TransformerslatestFinBERT pipeline for financial NLP sentimentmerquan_nawa_v2.py (E7 lazy load)
CCXT4.5.46Unified async crypto exchange API — 5 exchangesmain.py crypto price loop
REPORT LLMs & AI INFERENCE CONFIDENTIAL
ProviderModel(s)Endpoint / BaseRoleFallback Order
Anthropic Claudeclaude-sonnet-4-6 / haiku-4-5 / opus-4-7api.anthropic.comCommentary, CLAW assistant, trading bias1 — Primary
DeepSeek Officialdeepseek-v3, deepseek-v4-flash, v4-proapi.deepseek.comFast reasoning, analysis, scalp commentary2 — Fast LLM
Novita AIdeepseek-v3 (proxy)api.novita.ai/openaiDeepSeek relay when official API unavailable3 — Relay
NVIDIA NIM Cloudllama-4-maverick-17b, nemotron-70b, llama-3.3-70b, llama-3.1-405bintegrate.api.nvidia.com/v1Institutional inference, bias gen, CLAW models4 — Fallback
NVIDIA NIM Localllama-3.3-nemotron-super-49b-v1.5SaladCloud (deploying)Self-hosted 49B reasoning modelNew — Deploying
FinBERTProsusAI/finbertHuggingFace / local loadNLP sentiment on Finnhub headlines (NAWA E7)Always-on (lazy load)
MeraiMerqintel V1.0Github / local loadCommentary, CLAW assistant, trading biasAlways-on
REPORT GPU FRAMEWORKS CONFIDENTIAL
FrameworkFeatureUsed ForWhereStatus
JAX + XLAjax.jitJIT-compile options pricing + volatility modelsmerquanjaxsalad / SaladCloudActive
JAXjax.vmapVectorise full strike chains in single GPU callengines/options.pyActive
JAXjax.lax.scanGARCH(1,1) loop — GPU-native, zero Python overheadengines/har_rv.pyActive
JAXjax.gradExact Black-76 Greeks (delta/vega/theta)engines/options.pyActive
EquinoxJAX-native neural net layers (GRU, Linear, etc.)merquanjaxsalad imageIn Image
OptaxJAX gradient optimisers (Adam, etc.)merquanjaxsalad imageIn Image
BlackJAXBayesian inference on GPUmerquanjaxsalad imageIn Image
DiffraxDifferentiable ODEs/SDEs on GPUmerquanjaxsalad imageIn Image
PennyLane GPUlightning.gpuQuantum ML circuits (NAWA quantum engines)merquan_quantum_stack.pyConditional
cuPyGPU array ops + GARCH volatilitymerquan_quantum_stack.pyConditional
cuQuantumQuantum circuit simulation, HMM regimesmerquan_quantum_stack.pyConditional
PyTorch 2.10torch.jitDeep learning backbone + Triton kernel fusionmerqintel enginesActive
WebGPU / WebGL2Browser-side heatmap + footprint renderingqce_engine.html / heatmaptest.htmlActive
REPORT MARKET DATA SOURCES CONFIDENTIAL
SourceData TypeInstrumentsKey / CredentialRate Limit / Notes
Rithmic R|ProtocolLive ticks, DOM, BBO, depth-by-order, EOD327 CME/COMEX contracts (GCM6, 6E, 6J etc.)/etc/merquan/rithmic.envProtobuf/gRPC WSS — merquan-rithmic.service
Databento16yr historical OHLCVGC (gold futures) — 10M 1min barsdatabento 0.75.0 installed466K bars in QuestDB, 549K 1h bars
FinnhubQuotes, headlines, sentiment, tick dataStocks, ETFs, FX, metals3 keys: paid/ticks/free150 req/min — FinnhubLimiter token bucket
OANDA StreamingLive FX + metals tick streamXAU, XAG, XPT, EUR, JPY, GBP + 40+ pairsOANDA_API_KEY (env)tick_ingestor.py persistent stream
FREDMacroeconomic — Fed Funds rateFEDFUNDS seriesFRED_API_KEY (env)24h cache
CME GroupFutures options chains, market data6E (EUR), 6J (JPY), GC (Gold) options/etc/merquan/cme_merquan.envOAuth2 client_credentials
Bloomberg FIGIInstrument identifier mappingGlobal securitiesfigi_instruments.pyREST API
Binance / OKX / Bybit / Kraken / CoinbaseCrypto prices + arb spreads40+ tokensCCXT async15s refresh, cross-exchange arb detection
TradingView ScannerCME/COMEX/NYMEX futures OHLCVAll futures contractsPublic API (no auth)scanner.tradingview.com — 15s cache
REPORT OBSERVABILITY & TRACING CONFIDENTIAL
ToolVersionWhat It TracksAlert ThresholdsStatus
W&B Weave0.52.37All LLM calls (Claude/DeepSeek/NIM) — cost + latency + error rate>5% error in 5min / >$5/hr spendActive — 41+ run logs
wandb0.26.1ML experiment tracking, model runswandb.alert() on anomaliesActive
Sentry SDK2.58.0Application error tracking + crash reportingWired (verify active)
OpenTelemetry1.40.0Distributed tracing across servicesPackages present
journalctl/systemdService health logs for all 10+ systemd unitsActive