Skip to content

tharune/senthos

BranchesTags

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
README.md
README.md
 
 

Repository files navigation

Senthos (SCBC Hackathon 2026)

Senthos (SCBC Hackathon 2026): Polymarket-backed structured products on Solana devnet—NLP-filtered curated markets, bundle correlation and CVaR gates before persistence, Next.js 16 + Express + Supabase, Anchor (senthos_vault, senthos_ppn, senthos_lending), and the trained bundle under senthos-correlation-deliverables/ as read by the TypeScript scorer. Full route and program list: ARCHITECTURE.md in the application codebase.

Table of contents

  1. Overview
  2. Design principles
  3. Architecture
  4. Runtime topology
  5. Correlation model and artifacts
  6. Runtime integration (TypeScript)
  7. On-chain programs
  8. Backend and data plane
  9. Frontend
  10. Configuration and secrets
  11. Testing and verification
  12. Technical stack
  13. Cloud infrastructure (Akash Network)
  14. Repository layout
  15. Quick start
  16. Reference documents

Overview

Combined hackathon tree: Polymarket → structured products (baskets, tranches, PPN, lending surfaces), five-stage NLP + quality filter on GET /api/markets/curated, bundle correlation and CVaR gates on POST /api/bundles. Stack: Next.js 16, Express, Supabase, Anchor on devnet, ML artifacts in senthos-correlation-deliverables/. Below: boundaries between offline training (Python) and request-path scoring (TypeScript), plus Akash lease numbers for the GPU training run.

Design principles

  • Curated markets: GET /api/markets/curated is the single ingress for products that consume listings (MARKET_FILTER.md).
  • Correlation gate: POST /api/bundles applies weight optimization and tail-risk checks against audited envelopes (MODEL_INTEGRATION.md).
  • Training vs serving: trained model in the .tar.zst under senthos-correlation-deliverables/; request path uses the TypeScript heuristic in correlation.ts (see MODEL_INTEGRATION.md), no Python at runtime.
  • State placement: native baskets and PPN state live on-chain; tranche and lending mechanics are implemented at the service layer within hackathon scope, with an on-chain roadmap (PRODUCTS.md).

Architecture

Five layers:

┌──────────────────────────────────────────────────────────────────────────────┐
│  Client: Next.js 16 (App Router), wallet adapter, Solana JSON-RPC             │
├──────────────────────────────────────────────────────────────────────────────┤
│  API: Express — markets, bundles, hedge, PPN, lending, ML manifest, chain   │
├──────────────────────────────────────────────────────────────────────────────┤
│  Chain: Anchor — senthos: vault, PPN, lending (devnet)                        │
├──────────────────────────────────────────────────────────────────────────────┤
│  Data: Supabase; Polymarket Gamma; Anthropic API (filter pipeline)          │
├──────────────────────────────────────────────────────────────────────────────┤
│  ML: senthos-correlation-deliverables — audit, checksums, training bundle     │
└──────────────────────────────────────────────────────────────────────────────┘

Runtime topology

Standard local development uses three listening ports:

Process Port Role
Next.js (npm run dev) 3000 Public site / and authenticated /app/*
Express (npm run dev in backend/) 3001 REST API, Polymarket access, bundle creation, Solana transaction assembly
Monitor (same Node process as backend) 3002 Process monitor JSON (/data)

Scheduled price refresh runs inside the backend (ARCHITECTURE.md, §3).

Correlation model and artifacts

Objective: binary classification on pairs of prediction-market contracts: whether absolute return correlation reaches at least 0.6 (abs_corr_target >= 0.6), with decision threshold 0.70.

Reported metrics (senthos-correlation-deliverables/model_metrics_step12.json):

Quantity Value
Labeled rows 410,777
Train / validation split 328,623 / 82,154
Engineered features 20
Precision 0.9432
Recall 0.8831
ROC-AUC 0.9326
Regression RMSE 0.0836

Walk-forward summary (final_summary_step19.json): mean lift +4.82% versus a naive baseline; p = 3.10 × 10⁻⁵; audit checklist completed.

Monte Carlo (monte_carlo_step14.json): 50-leg basket, 20,000 paths, 7-day horizon, σ_daily = 4%; VaR and CVaR recorded for the audited risk envelope.

Deliverables directory (senthos-correlation-deliverables/): step-level metrics (JSON/Markdown), SHA256SUMS.txt, manifests, and senthos-correlation-production-20260417T071458Z.tar.zst (Python training bundle). Paths embedded in the archive reference /root/senthos-correlation/... on the training host.

Runtime integration (TypeScript)

Implementation: backend/src/services/correlation.ts.

Export Role
loadArtifacts() Lazy, memoized load of audit JSON; exposes thresholds, Monte Carlo envelope, walk-forward flags
scoreLegPair(a, b) Noisy-OR combination of text, tag, and temporal similarity signals (MODEL_INTEGRATION.md, §3.2)
optimizeWeights, assessBasketRisk, getModelManifest Basket weights, CVaR ratio gate, manifest for /api/ml/*

The HTTP tier does not invoke Python during request handling.

On-chain programs

Program Function
senthos_vault Native basket vault: deposit, redeem, resolve/finalize, fees
senthos_ppn Principal-protected notes; mock yield adapter
senthos_lending Lending pool: supply, borrow, repay, withdraw

Program IDs, PDAs, and instruction matrices: ONCHAIN.md, ARCHITECTURE.md, and programs/*.

Backend and data plane

  • Framework: Express 4, TypeScript, Zod, @coral-xyz/anchor, Supabase client, express-rate-limit, node-cron.
  • HTTP surface: backend/src/routes/ — bundles, markets, hedge, PPN, lending, alerts, diagnostics, on-chain status, ML manifest. Enumeration: ARCHITECTURE.md, §§5–6.
  • Services: backend/src/services/ — correlation, NLP/market filter, pricing, Polymarket adapters, Solana bridge.

Frontend

  • Stack: Next.js 16.2.4, React 19, Turbopack (dev), Tailwind CSS v4.
  • Wallet: @solana/wallet-adapter-*; shell under app/app/.
  • Routes: marketing root app/page.tsx; product routes under app/app/* (ARCHITECTURE.md, §5).

Configuration and secrets

  • Root: .env.local (templates: .env.local.testnet.example, devnet variant). Cluster switching: SWITCH-CLUSTER.command.
  • Backend: backend/.env from .env.testnet.example or .env.devnet.example — Supabase, RPC endpoints, program IDs, optional DISABLE_RATE_LIMIT, Anthropic credentials for the filter stages.
  • Variable index: CREDENTIALS.md.

Testing and verification

  • Programs: tests/ — Anchor integration (Mocha/Chai).
  • API / NLP: backend/src/scripts/ — e.g. test-api.ts, nlp-probe.ts (26 checks per ARCHITECTURE.md).
  • Liveness: GET /api/health, GET /api/onchain/status, GET http://localhost:3002/data.

Technical stack

Layer Technology
Frontend Next.js 16.2.4, React 19.2.4, TypeScript 5, Tailwind 4
Backend Express 4.21, TypeScript 5.9, tsx
Chain @solana/web3.js, @coral-xyz/anchor 0.30.1, SPL Token
Programs Rust, Anchor (senthos_vault, senthos_ppn, senthos_lending)
Persistence / feeds Supabase JS v2, Polymarket Gamma API
NLP Anthropic SDK
ML (offline) Python, scikit-learn, LightGBM
ML (online) TypeScript correlation service (heuristic + frozen audit metrics)
Training compute Akash Network — GPU lease for correlation pipeline

Cloud infrastructure (Akash Network)

Correlation training (features, LightGBM/sklearn on 410,777 labeled rows per audit JSON, walk-forward, Monte Carlo) ran on Akash Network: SDL v2 lease, settlement in AKT.

Training used only the leased GPU node; repo artifacts under senthos-correlation-deliverables/ were built on that host (/root/senthos-correlation/... paths inside the tarball).

Deployment (training phase)

Phase Nodes Workload Wall-clock duration
Training 1 × GPU Python correlation pipeline and audit scripts ~9.5 h

Resource profile (hackathon lease)

Declared resources for the hackathon lease (confirm against your saved SDL / provider console).

Resource Declared value
CPU 32.0 cpu.units (32 threads under SDL v2 resource semantics)
Memory 100 GB
Storage 400 GiB
GPU 1× NVIDIA H100 (vendor.nvidia.model: h100)
Replicas 1

Cost

Line item USD Basis
GPU training lease 34 Lease-related AKT for this workload, converted to U.S. dollars at hackathon close-out (rounded). Wallet or block explorer history is the reconciliation source.

Settlement in AKT via the Akash marketplace. The amount covers the ~9.5 h training window; it excludes unrelated leases, local hardware, and third-party API egress.

SDL excerpt

Same shape as the training deployment (adjust image, env, bids to match your lease).

---
version: "2.0"

services:
  senthos-train:
    image: your-registry/senthos-correlation-train:latest
    env:
      - RUN_ID=senthos-2026-correlation
    expose:
      - port: 8080
        as: 80
        to:
          - global: false

profiles:
  compute:
    trainer:
      resources:
        cpu:
          units: 32.0
        memory:
          size: 100Gi
        storage:
          - size: 400Gi
        gpu:
          units: 1
          attributes:
            vendor:
              nvidia:
                - model: h100
  placement:
    akash:
      pricing:
        trainer:
          denom: uakt
          amount: 10000

deployment:
  senthos-train:
    akash:
      profile: trainer
      count: 1

Pin images by digest in production; keep keys out of SDL; amount is a bid cap — settled rate is whatever the provider accepted.

Repository layout

High-level mirror of the combined tree (paths relative to that codebase):

app/                       # Next.js App Router
backend/src/               # Express routes, services, Solana, IDL
programs/                  # senthos_vault, senthos_ppn, senthos_lending
senthos-correlation-deliverables/
scripts/  migrations/  tests/
ARCHITECTURE.md, MODEL_INTEGRATION.md, MARKET_FILTER.md, …

Quick start

From the application repository root (clone/runbook internal to your team):

cp .env.local.testnet.example .env.local
cp backend/.env.testnet.example backend/.env
npm install && (cd backend && npm install)
# Terminal 1: (cd backend && npm run dev)
# Terminal 2: npm run dev

Devnet: use SWITCH-CLUSTER.command and the env templates documented beside it in that tree.

Reference documents

File Contents
ARCHITECTURE.md Topology, routes, mathematics, change recipes
MODEL_INTEGRATION.md Model contract and HTTP behavior
MARKET_FILTER.md NLP filter stages
PRODUCTS.md Product taxonomy and roadmap
ONCHAIN.md, ONCHAIN_DESIGN.md Program design
CREDENTIALS.md Environment variables
AUDIT.md, SECURITY.md Security notes
senthos-correlation-deliverables/README.md Artifact inventory

About

Senthos (SCBC Hackathon 2026) — public architecture & methodology; source in SCBC repo.

Resources

Readme
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

 
 
 

Contributors