Method · How Synaptech builds

AI drafts. Senior engineers decide.

We build with the same AI tools many founders start with, and we use them hard: scaffolding, first drafts, tests and docs. Then a senior engineer makes every call that is expensive to reverse. It is the same rule we build into the AI features we ship for clients.

Get your free Production Gap Review or just book the call

30 minutes. A written gap map. No pitch unless you ask for one.

Product live but empty? Get a free Growth Engine Audit

AI drafts

Scaffolding
First drafts
Tests
Docs

Senior engineer review

Engineers decide

Architecture
Security
Money paths
Releases

Fig. 02 · The division of laborEvery change reviewed · every release owned

AI speeds us up

Scaffolding
Project skeletons, boilerplate and configuration. We start where you started: some of our own repositories began as scaffolds in Bolt.new, an AI app builder.
First drafts
Features, migrations and refactors drafted fast, then read diff by diff by the engineer who owns the change.
Tests
Test scaffolds and edge-case lists drafted from the spec, and kept only when they prove something.
Docs
Runbooks, READMEs and handover notes drafted from the code, then checked against it.

of our repos started as scaffolds in Bolt.new, an AI app builder, before our engineers took over

receipt

SOURCE · Synaptech portfolio repositories

METHOD · .bolt project directories found during the September 2026 portfolio audit, counted conservatively: four web front-ends and two product prototypes. More of our repositories carry .bolt folders and are not counted here.

VERIFIED · 2026-09

CLASS · repository

✓ every number on this site opens its receipt

CLAIM · bolt-scaffolds

How we count

Senior engineers decide

Architecture
Data model, tenancy and service boundaries. Decisions that are expensive to reverse are made by a senior engineer, in writing.
Security
Auth, roles, secrets, and exactly what the AI features in your product may read, and never write.
Money paths
Payments, refunds, ledgers and reconciliation. Deterministic code with tests to the penny; no model does the arithmetic.
Releases
What ships and when, with a rollback plan. Every change that ships is reviewed by a named senior engineer who is accountable for it.

We hold the same line inside the AI features we build for clients: the model proposes, code decides.

See the six guardrails we ship

How much faster does this make us? We publish one documented delivery velocity example, with its receipts, and we tell it once rather than turning it into a rate card.

See the documented delivery velocity

The LLM extracts. The server decides.
Model output is a proposal. Deterministic code resolves it against real records, and code does any writing.
Runs in SkillSchedule’s session setup · a 12-rooftop auto group’s statement pipeline
Read-only data roles
The database role an AI feature runs under cannot write. The database enforces it, and tests prove it.
Runs in a 12-rooftop auto group’s accounting assistant · Pour Path’s AI data API
Audit logs
Sensitive actions leave a row that says who did what, and why.
Runs in Permission overrides with a stated reason at a 12-rooftop auto group · every SMS, email and call attempt at TriaPet
Kill-switches
Capabilities that touch money or message customers ship disarmed, behind an explicit switch.
Runs in Willow’s money and messaging features
Penny-exact validation
A statement must balance against its printed totals to the penny, or it is quarantined for review.
Runs in a 12-rooftop auto group’s bank-statement pipeline
Disclosure boundaries
Tested limits on what an AI may say out loud, with hard-fail gates on anything outside them.
Runs in TriaPet’s voice agent

Case files: the rules in production

SkillSchedule's "Your Assistant Coach" feature grid on blue: auto payments, smart reminders, QR family connect and AI session setup.

Case file 01 · Scheduling assistant

The LLM extracts. The server decides.

ClientSkillSchedule

Coaches create sessions in plain language. The model proposes names, dates and times; deterministic server code resolves them against the coach’s real roster. The model has no write access at all.

A dual-provider assistant: GPT-4o first, Claude as automatic fallback
Entity resolution against the real roster, in code
No write path for the model

Read the SkillSchedule case

Overhead flat-lay of sealed Craveble meal trays, including chicken parmesan, chile relleno and slow-smoked pulled pork, set diagonally on powder blue.

Case file 02 · AI CFO

The AI can look. It cannot touch.

Venture we co-operateCraveble

Bank and card accounts sync through Plaid into Postgres. Claude classifies each merchant into a real chart of accounts; code, not the model, decides P&L grouping and tax deductibility. The CFO advisor answers from a live snapshot of the books.

The AI classifies; deterministic code decides the accounting treatment
Transfers and card payments de-duplicated, so burn is never double-counted
An idempotent sync that survives a failed page of results

Read the Craveble case

QuestionWhich entities paid the most in bank fees this quarter?

-- runs under a read-only role
-- shown with every answer
SELECT entity, SUM(fee_amount)
  FROM statement_lines
 WHERE period = :quarter
 GROUP BY entity;

SELECT-only role · writes rejected by the SQL guard

Case file 03 · Accounting assistant

Read-only, and penny-exact.

Clienta 12-rooftop auto group

Ownership asks questions in plain English. Answers come through a SELECT-only database role behind a two-layer SQL guard, and every answer shows the SQL it ran. Underneath, the model only handles the genuinely fuzzy part of each statement and must agree with the amounts extracted deterministically.

Tests prove the SQL guard rejects writes
The SQL behind every answer, visible to the person asking
Statements that don’t balance are held for review, never silently included

10 accounts · 9 entities

bank statements reconciled to the penny, every one checked against its printed totals

receipt

SOURCE · Finance pipeline repository for a 12-rooftop auto group

METHOD · Validation logic and database re-verification in the pipeline; failures are quarantined, never silently included.

VERIFIED · 2026-09

CLASS · repository

✓ every number on this site opens its receipt

CLAIM · penny-reconciliation

How we count

Read the platform case

TriaPet homepage on a phone: the headline, the licence-verification form and the start of a spoken case card.

Case file 04 · Voice agent

Disclosed, bounded, graded.

Venture we operateTriaPet

When a text to a clinic goes unanswered by a case’s deadline, Tria, a Retell.ai voice agent, calls the front desk within calling hours. It gives a versioned spoken disclosure, asks exactly one question and posts the answer to the sender’s live status board.

A tested disclosure boundary: no owner names, prices or clinical detail
Simulated calls graded against a written rubric, with hard-fail gates for impersonation, invented detail and identity leaks
Every SMS, email and call attempt leaves an auditable ledger row

Read the TriaPet case

Is AI writing my product?

AI drafts parts of it: scaffolding, first drafts, tests and docs. A senior engineer decides everything that ships and is accountable for it by name. Architecture, security, money paths and releases are decided by people.

Are the AI features you build safe to put in front of customers?

They are built so the model can’t do the dangerous part. The LLM extracts and the server decides; data roles are read-only; sensitive actions leave an audit trail; anything that touches money or messages customers ships behind a kill-switch; and anything that speaks for you has a tested disclosure boundary.

Does AI make the project cheaper?

It makes parts of the work faster, which is why the math can work at your size. We don’t publish blanket discount percentages. We publish one documented delivery velocity example on our homepage, with its receipts, and the free review ends with a fixed, scoped price.

Which AI models do you use?

Whichever fits the job, often more than one. Products we have shipped run Claude, Gemini and GPT-4o, some with an automatic fallback from one provider to another. The model is a component. The rules around it are the product.

05 · The front door

Find out what’s between you and production.

Bring your prototype, or just the idea. Leave with a written gap map — deployment, security, payments, integrations, growth readiness — before you spend a dollar. If we’re not the right team to close the gaps, the map is still yours.

Get your free Production Gap Review or just book the call

30 minutes. A written gap map. No pitch unless you ask for one.

info@synaptech.io +1 (925) 255-9507 Pleasanton, CA

AI drafts. Senior engineers decide.

One path, from idea or prototype to customers.

Where AI speeds us up. Where people decide.

Scaffolding

First drafts

Tests

Docs

Architecture

Security

Money paths

Releases

The AI can look. It cannot touch.

The LLM extracts. The server decides.

Read-only data roles

Audit logs

Kill-switches

Penny-exact validation

Disclosure boundaries

Case files: the rules in production

The LLM extracts. The server decides.

The AI can look. It cannot touch.

Read-only, and penny-exact.

Disclosed, bounded, graded.

Questions about how we use AI.

Find out what’s between you and production.