Services Ventures Work About Start a project

Our work

The Work

Nearly every project here runs in production, and the public ones are linked.

Platform & Open SourceOpen source2026

SPECTER

An open-source agent for getting cited by AI search

Search engines rank pages, and generative engines cite sources. SPECTER is a skill graph that teaches a coding agent the whole loop: watch which sources Perplexity and ChatGPT pull from, collect the intelligence, find the gaps, write the answer, publish it, and check whether the citation showed up.

skill-graph nodes: 43
capability clusters: 5
license: MIT
monitoring cadence: Weekly

Claude Code pluginSkill graphx402Headless CLIlaunchd

Read the write-up Read the source

Interface & ArtLive2026

THRESHOLD

Eight minutes tracing how perception begins

THRESHOLD is a six-stage progressive web app that carries a viewer from in-utero sensation to the first social smile. It is written against raw WebGL2 and the Web Audio API, with no rendering library, no skip controls, and nothing that leaves the device. It is an art piece, and those constraints are part of the work.

stages: 6
minutes, unskippable: 8
rendering libraries: 0
bytes leaving the device: 0

WebGL2Web Audio APIMediaPipeNext.jsIndexedDB

Read the write-up Open the piece

Platform & Open SourceOpen source2026

GHG Calculator

One emissions engine, three interfaces, no API keys

This GHG Protocol Corporate Standard calculator covers all three scopes and bundles 967 emission factors from six peer-reviewed databases inside the package. It runs as a command-line tool, as an MCP server any AI agent can call, and as a skill a coding agent loads on demand.

embedded emission factors: 967
peer-reviewed source databases: 6
scopes covered: 3
API keys required: 0

PythonuvMCPClaude Code skillGHG Protocol

Read the write-up Read the source

Platform & Open Source2026

The Symbai Studios Platform

A monorepo where the rules enforce themselves

Every capability we ship (customer records, forms, reporting, integrations, AI) is a package that any venture can pull in. The apps themselves are thin deployment shells. The architecture runs four layers deep with dependencies flowing downward only, and seven automated gates keep it that way.

TypeScript packages: 25
deployed applications: 16
automated architecture gates: 7
ventures on one codebase: 3

NxpnpmTypeScriptPrismaPostgreSQL

Read the write-up

Agentic Systems2026

Agent Infrastructure

Scheduling, supervision, and memory for agents that run unattended

Agent Infrastructure is a set of scheduled jobs that keeps an agent practice honest. One digests every coding session into a searchable knowledge base, another watches every other job and complains when it goes stale, and a third runs a weekly citation audit. Together they let the rest of the work run without supervision.

agent CLIs, one pipeline: 4
tests: 47
health monitoring: Hourly
unattended operation: 24 / 7

PythonlaunchdClaude CodeCodexGrok

Read the write-up

Applied AI2026

Regen Explorer

Deterministic chain reads first, AI surfaces second

A reference instance for Regen Network's ecological credit registry, split into two tiers. The first reads the chain deterministically with no key and no model in the path. The second adds AI surfaces that answer questions about the registry, find grants, and review projects.

credit classes read live: 13
projects: 60
credit batches: 80
tiers, one behind a gate: 2

Next.js 16Anthropic MCP connectorUpstashZodCosmos REST

Read the write-up

How this gets built

One Specification, Many Implementers

Write the specification once

A task is not ready for a model until its acceptance criteria are written down and testable. The specification is the part that lasts. Which model types the code is a detail we change freely, often in the middle of a project.

Route the work to the right lane

Routine implementation goes to a cheaper model from a different vendor, while architecture and commitment decisions go to the most capable model available. Running several model families against the same specification surfaces disagreements that a single model would have smoothed over.

Make the rules mechanical

Architecture gates, type strictness, coverage thresholds, and fabrication guards run at commit time and in continuous integration. An agent will not remember a convention across a long session, so the convention has to reject the commit itself.

Ship on evidence

Looking finished is never enough to ship. An enriched contact record gets thrown out unless it comes with a source URL, and a build that fails its checks never reaches production. The default answer is no, and our job is to produce enough evidence to change it.

Supervise what runs unattended

Scheduled agents fail silently, so a monitor watches log freshness and process health across every always-on job and reports only when the failure set changes. We digest every session into a searchable record, so the reasoning behind a decision outlives the session that made it.

Get in touch

Start a Project

Tell us about your business and where you want to take it next, whether that is the website, the offering, or the platform behind it.

Book a call Get an estimate

hello@symbai.studio