Synroi

Stop paying twice for the
same AI computation.

Starting with SQL.

Claude GPT-4o Gemini
╲ | ╱
Same Computation
One Execution
Cache Reuse

AI agents often express the same computation in many different ways. Synroi identifies equivalent computations before execution.

Four steps. One execution.

1

Receive

Multiple SQL queries arrive from different LLMs, agents, or users.

2

Identify

Synroi recognizes when queries compute the same result — even when they look completely different.

3

Execute once

The first query runs against the database. Every equivalent query after it does not.

4

Serve from cache

Subsequent equivalent queries return instantly from semantic cache.

What is it. Why buy it. Why trust it.

What is it?

Computation Identity

The technology

Different expressions of the same computation receive the same identity. This is not string matching — it is structural analysis of what the computation does.

  • Deterministic computation identity
  • Conservative equivalence detection
  • Consistent across servers, time, and deployments
  • Conservative by design — only proven-safe equivalences
Why buy it?

Enterprise Gateway

The product

A drop-in gateway that sits between your applications and your database. Semantic caching, deduplication, and audit — with no code changes required.

  • PostgreSQL, Snowflake, BigQuery, Databricks, Redshift
  • Shared identity across SQL dialects
  • Minimal latency overhead
  • Complete audit trail for every query
Why trust it?

Privacy

The differentiator

Synroi runs entirely inside your infrastructure. Offline by default. No outbound connections. Your queries, your schemas, your data — they never leave.

  • Zero outbound network connections
  • No data collection, no license server, no phone-home
  • Customer owns all data and identities
  • Federated Insights is always optional

Three LLMs. Three queries. One computation.

These are not merely formatted differently. The predicates are written in different orders and some comparisons are reversed. Synroi identifies them as the same computation.

Input queries

Claude
GPT-4o
Gemini

Result

Claude GPT-4o Gemini
Same computation identity
waiting...
1
execution
2
cache hits
Query 1 (Claude)-
Query 2 (GPT-4o)-
Query 3 (Gemini)-
Rows returned-
-
Queries
-
Unique
-
Cache hits
-
Avoided
-
Reduction
-
Status

Try any SQL query.

Enter a query, execute it, then try a variation. Same computation? Same identity. Served from cache.

Results appear here.

Tested against 1,635 AI-generated queries.

5 frontier LLMs were given 327 identical analytics tasks. Every valid query was processed. Every malformed output was safely rejected.

1,633
Valid SQL processed
5
Frontier LLMs
62.3%
Executions avoided
1,017
Redundant queries caught

The 2 malformed queries were hallucinated LLM outputs containing invalid syntax — not valid SQL. Synroi correctly processes all valid SQL and safely rejects malformed AI output.

Every avoided execution saves money.

Modelled scenario: At 100K queries/day on Snowflake with AI agent workflows, the model estimates up to $443K/year in savings from eliminated duplicate queries and reduced warehouse compute.
Modelled estimate under stated assumptions. Not observed customer savings.

Three ways to deploy. All run inside your infrastructure.

Every deployment option runs offline by default. No data leaves your environment.

Docker

Evaluate in minutes

Single container with SQLite backend. No dependencies. Run it, send a query, see identity and caching in action.

docker compose up -d curl localhost:3400/health
  • One command to start
  • SQLite or PostgreSQL backend
  • Dev mode for instant evaluation
  • Same binary as production

Helm Chart

Production Kubernetes

Full Kubernetes deployment with PostgreSQL, NetworkPolicy, PodDisruptionBudget, and Grafana dashboards included.

helm install synroi ./charts/synroi \ --set database.host=postgres \ --set license.existingSecret=synroi-license
  • Pod Security Standards (restricted)
  • Egress locked to database only
  • Prometheus metrics built in
  • Grafana dashboard auto-provisioning
  • OpenShift compatible

Air-Gapped

Maximum isolation

Offline license verification. Zero outbound connections. Verify with tcpdump. Designed for regulated environments.

docker run --net=none \ -e CCO_LICENSE_FILE=/license.json \ -v ./license.json:/license.json:ro \ synroi:latest
  • Ed25519 offline license verification
  • No DNS, no HTTPS, no phone-home
  • Read-only root filesystem
  • Runs as non-root (uid 65534)
  • Cluster binding for enterprise
License tiers

Standard

Single dialect, core capabilities
  • PostgreSQL dialect
  • Canonicalization + semantic cache
  • API authentication
  • Audit trail

Professional

Multi-dialect, full observability
  • PostgreSQL + Snowflake + selected dialects
  • Prometheus metrics
  • Grafana dashboards
  • Batch canonicalization API
  • SQL Analyze tool

Enterprise

All dialects, maximum control
  • All 5 dialects
  • Cross-dialect computation identity
  • Cluster binding
  • Federated Insights (optional)
  • Air-gapped deployment support
Try Analyze free

Runs entirely inside your infrastructure.

Offline by default. Zero outbound connections. Your data never leaves.

No license server. No data collection. No phone-home. If you later choose to participate in Federated Insights for anonymous industry benchmarks, that is entirely optional and exists to benefit your deployment.

Offline by default

The gateway makes zero outbound network connections. No DNS lookups. No HTTPS calls. No reporting endpoints. Verify with tcpdump.

You own everything

Your SQL, your schemas, your results, your cache, your computation identities. Synroi owns the software and algorithms. That's it.

Conservative by design

Only proven-safe equivalences are applied. When in doubt, Synroi treats queries as different. Safety over aggressiveness.

Auditable

Run with --net=none. Inspect the binary dependencies. strace the process. Every claim is verifiable.

Architectural Invariants
I-1Customer SQL never leaves the customer environment.
I-2Customer schemas never leave the customer environment.
I-3Customer results never leave the customer environment.
I-4Customer caches never leave the customer environment.
I-5Offline deployment is the reference deployment.
I-6Federated Insights is always optional.
I-7Customer value precedes Synroi value.
Technical details
Synroi normalizes SQL queries into a canonical computation representation and computes a deterministic identity from the result. Queries that are structurally equivalent — differing only in formatting, predicate ordering, comparison direction, or other syntactic variations — are normalized to the same representation and receive the same identity.

The normalization engine applies a set of conservative equivalence rules. Each rule is provably safe: it will never identify two queries as equivalent if they could produce different results on any valid dataset.

Identity is deterministic. The same query always produces the same identity, regardless of which server, dialect, or deployment processes it.

Five SQL dialects are supported: PostgreSQL, Snowflake, BigQuery, Databricks, and Redshift. Each dialect is parsed independently and normalized to the shared canonical representation, enabling cross-dialect computation identity.

The gateway includes full provenance tracking — every computation identity can be traced to the specification and process that produced it. Built in Rust.