WATAM DeFi Intelligence collects a token's swaps directly from the blockchain, checks that the data is complete, and shows what happened pool by pool — with unusual activity flagged and the evidence attached. It runs on infrastructure you control, and every report states exactly which pools and hours it covers.
Your infrastructure. Your blockchain data. Verifiable intelligence. Deployed per customer — no shared hosted service — and pilots are scoped individually.
Real screens from the self-hosted web application: the same VELO finding as the app lays it out in a desktop, tablet (820 px) and phone (390 px) browser. Behind them, the real hourly VELO volume on OP Mainnet for 15–16 Oct 2026, with the frozen baseline dashed and the two flagged spikes marked. Historical snapshot computed from swap events collected over public JSON-RPC — not live. It is a responsive web app, not a native app.
Swaps are read from public JSON-RPC or your own endpoints, proven complete for every block range and cross-checked before anything is published. A protocol family is supported only after its decoded swaps matched transaction receipts exactly on several networks. Uniswap v4 is verified differently: its pools share one contract and can carry hooks, so each pool is verified on its own, and only individually verified pools are ingested.
Every figure carries its coverage, its valuation method and the SQL behind it, so analysis can be reproduced, reviewed and shared.
See how your token trades across chains and pools — volume, liquidity and valuation coverage — and get flagged when activity departs from its baseline.
Source-backed token and protocol research with stated coverage, so a conclusion can be traced back to the swaps that produced it.
Anomaly findings for investigation triage: statistical observations with baselines, limitations and alternative explanations, never accusations.
Client-ready reports from verified data on private infrastructure, with an AI analyst that answers in several languages without inventing numbers.
Three stages, each with its own checks. Nothing reaches a report or an AI answer without passing the stage before it.
Ingestion runs prove block coverage, confirm empty answers with a second provider and publish to ClickHouse only after quality gates pass. USD comes from stablecoin legs or an hourly price feed configured for the installation; unvalued swaps are never counted as $0.
Covered-hour metrics, window comparisons and an anomaly engine with frozen baselines. Reports state exactly which pools and hours they cover.
The AI phrases answers from the verified facts of each question. Every number must match a fact the sentence cites; anything unproven is replaced by a clearly labelled deterministic answer.
The workflow end to end: verified data, a token report, an anomaly finding with its SQL evidence, a two-window comparison and the analyst.
The narration is an AI-generated voice. The screens come from a 20 Oct 2026 build with historical data. Captions are on by default.
Three analytical tasks the current release supports on its verified networks and protocol families. Each one states what it covers and what it does not.
Token teams, protocol analytics, research desks
Paste a token contract. The application checks it on-chain, finds its pools and says which ones it can decode. It then collects the missing hours through the same gated pipeline and reports swaps, token volume, buyers and sellers, and routing per pool — with the covered hours stated.
Covers the selected pools on supported networks, not the token's whole market.
Risk, market-integrity and research teams
The anomaly engine flags volume, activity and liquidity shifts against a baseline frozen at onset, plus concentration, imbalance and repetitive patterns. Every finding carries its facts, method, parameters, limitations and alternative explanations, so it can be checked before anyone acts on it.
A statistical observation, never a claim of intent or manipulation.
Growth, treasury and research teams
Compare a token's activity before and after a listing, an incentive change or an event. A comparison is marked comparable only when every pool is covered in both windows; otherwise the report says it is incomplete instead of showing a misleading change.
Historical windows that have been ingested; not live streaming.
Ask in English, Turkish or German and keep the conversation going: follow-up questions keep the token and topic. Each answer is labelled either as a validated AI answer or as a deterministic answer built from the same verified facts.
What happened to VELO volume on OP Mainnet on 16 October?
| Inference option | Status | Notes |
|---|---|---|
| Local models via Ollama | Tested | Runs on your hardware; a memory guard prevents overloading the host |
| Anthropic Claude API | Integrated | Optional and off by default; no paid call until an administrator authorises it within limits. Live calls not yet verified |
| OpenAI API | Integrated | Same controls as Claude. Live calls not yet verified |
| Self-hosted OpenAI-compatible servers | Integration option | vLLM, llama.cpp, LM Studio or an organisation gateway through the same adapter; GPU serving not yet tested |
The same frozen fact set (10 questions in English, Turkish and German; 8 answered by the model, 2 declined by deterministic pre-checks) and the same validators, on a 16 GB machine.
| Model | Accepted by validation | Grounding failures | Avg latency | Memory |
|---|---|---|---|---|
| gemma3:4b | 8 / 8 | 0 | 8.3 s | 2.9 GB |
| qwen3.5:9b | 6 / 8 | 2 | 22.4 s | 5.7 GB |
In release testing on 2026-10-17, gemma3:4b answers to an English question and English and Turkish follow-ups were all accepted. In a later run a Turkish answer that cited the wrong facts was rejected and replaced by the deterministic answer — exactly what the validation is for. One machine, not a performance promise.
Reviewed snapshots of historical windows from October 2026. Every figure here was recomputed before publication from swap events our own pipeline collected over public JSON-RPC. USD values are deliberately not shown: they depend on an external price feed. These examples illustrate the method; they are not live data and not a statement about any protocol.
14 Oct 2026, 12:00 UTC to 17 Oct 2026, 08:00 UTC · 4 ingested Uniswap v3-style pools · reviewed snapshot, not live
Every Swap event emitted by the selected pools in this window, counted hour by hour. These are selected-pool event measurements: not the total ARB market and not transaction-net trading volume.
| Pair | Protocol family and pool | Fee tier | Swap events |
|---|---|---|---|
| WETH / ARB | Uniswap v3-style 0xc6f7…396a | 0.05% | 40,956 |
| ARB / USDC | Uniswap v3-style 0xcda5…dad8 | 0.05% | 18,586 |
| ARB / USDC | Uniswap v3-style 0xaebd…ccc3 | 0.3% | 5,114 |
| WETH / ARB | Uniswap v3-style 0x92c6…8f00 | 0.3% | 107 |
| 4 pools | receipt-verified Uniswap v3-style decoder | 64,763 |
16 Oct 2026, 12:00 UTC to 17 Oct 2026, 00:00 UTC · 12 pools discovered and verified, 6 ingested · reviewed snapshot, not live
An example of what the anomaly engine reports — and of what the data does not cover. Market coverage is partial and the report says so. USD values are not shown (see above).
We separate what has been tested on real data from what is integrated, experimental or not supported. "Reference installation" means our own deployment; every customer installation verifies its own backups and recovery.
| Capability | Status |
|---|---|
| Ingestion, token explorer, anomaly engine, reports and window comparisons on Base, Arbitrum One, OP Mainnet and Ethereum | Tested on real data |
| Uniswap v3-style, Uniswap v2-style and Solidly-style decoders (decoded swaps matched transaction receipts exactly) | Tested on real data |
| Uniswap v4: six individually verified Ethereum pools (PoolId/PoolKey match, two-provider agreement, sampled receipt reconciliation) | Tested on real data |
| Access control: every API route requires sign-in by default, organisation roles, personal API tokens, single-use invitations for onboarding | Tested and in production use |
| Sandboxed query engine: file, network and extension access blocked at validation and at execution | Tested and in production use |
| Encrypted off-site backup to Cloudflare R2, restore-verified on a separate isolated server | Verified on our reference installation (Oct 2026) |
| Local models via Ollama, including a multilingual multi-turn conversation | Tested on one machine |
| Self-hosted Docker Compose installation: TLS proxy, first administrator, your own RPC endpoints, upgrade and rollback | Tested on Docker Desktop |
| Kubernetes deployment | Validated on a local single-node cluster |
| PancakeSwap pools, Solana research, one-click token analysis | Experimental |
| Anthropic Claude API and OpenAI API (optional, off by default, your own account) | Integrated, live calls not yet verified |
| Dune (optional, off by default, your own account) | Integrated, not verified live |
| vLLM and GPU inference servers | Not yet verified |
| Other Uniswap v4 pools, other networks, hooks that change swap amounts, native ETH | Not supported |
| Hosted multi-tenant SaaS | Not offered |
Only the TLS proxy is exposed. ClickHouse, pipelines and inference stay on internal networks, and secrets are encrypted at rest.
| Profile | Status | Target | Notes |
|---|---|---|---|
| Business server | Tested on Docker Desktop | Linux with Docker Compose, 32 GB RAM recommended, 50 GB+ persistent storage, optional NVIDIA GPU | Read-only preflight installer, first-administrator setup, upgrade and rollback by release tag. Not yet installed on a customer server |
| Local evaluation | Tested | A 16 GB machine | Small local models, memory guard, automatic ingestion off by default |
| Enterprise | Validated locally | Kubernetes, separate GPU inference servers | Kubernetes manifests with optional Airflow. Upgrade, rollback, persistence, network policies and backups validated on a local single-node cluster; production clusters, managed cloud, external TLS ingress and GPU inference nodes not yet validated |
Backups are encrypted before they leave the server, and the key stays with you. Point them at your own S3-compatible bucket for off-site copies; each backup can be restore-verified on a separate, isolated server without modifying production data. On our reference installation an encrypted backup to Cloudflare R2 and its isolated restore were verified end to end in October 2026; each customer installation runs its own backup and restore verification.
Starting points, not performance promises: capacity depends on the measured hardware.
Tell us which tokens, protocols and networks matter to you. The first step is a discovery conversation; if it fits, we agree a fixed, measurable pilot scope in writing before anything is installed.
Opens a short email template: organisation, contact, what you want to analyse, preferred deployment and AI. Or write to admin@wearetheartmakers.com
Your analytical questions; the chains, tokens, pools and data sources involved; deployment and data-handling constraints; integrations you need; and the current limitations that apply to your case.
A Linux server with Docker (32 GB RAM recommended, 50 GB+ storage), optionally your own RPC endpoints and a local or cloud AI model.
Confidential datasets, API credentials, private keys or passwords — not by email and not during the pilot.
Pilot scope, deployment feasibility, data-source permissions, licensing and commercial terms are agreed in writing. A first conversation is not a commitment on either side.
Stated plainly, so you know what the numbers cover.