Syltech AI Systems

What Syltech does

Across the full stack

Agentic AI

Governed agent loops with durable execution and spend guards.

Crash-safe, exactly-once execution with journal-based recovery, and hard spend and time ceilings on every agent loop.

AI platforms

Model and tool orchestration, internal platforms, secure delivery.

Model and tool orchestration with retrieval and memory, shipped through automated, secure delivery that teams can actually operate.

Identity & authz

RBAC and ReBAC, SSO, and policy across a multi-tenant hierarchy.

Hierarchical RBAC plus ReBAC on Ory Keto relation tuples, resolved at check time across a multi-tenant hierarchy, with enterprise SSO on Ory Kratos.

On-device & edge

Quantized LLMs running offline on phones and embedded hardware.

On-device LLM inference on Android with LiteRT, and object detection running on embedded Renesas processors.

Durable execution

Long-running workflows that survive crashes and resume exactly once.

Built rerun, a durable execution engine for exactly-once workflows, plus DBOS-backed pipelines with retry, reconciliation, and safe resumption across regions.

Model training

Training and evaluating models, not just prompting them.

Training and evaluating models end to end, from data to metrics, not just prompting hosted ones.

Retrieval (RAG)

Grounding LLMs in private data so answers stay cited.

GenAI summarization and classification grounded over 1M+ document-review workloads on a legal platform.

Mobile apps

Android apps with on-device inference in Jetpack Compose.

An Android tutor app that runs a quantized LLM fully on-device in Jetpack Compose, offline-first.

Cloud & Kubernetes

Private cloud and Kubernetes on OpenStack, provisioned as code.

A VMware-to-OpenStack migration driven by a self-service developer platform that replaced ticket-based provisioning.

agentic system · architecture

Approach

How Syltech builds

Secure by construction.
Operable by default.

A system should be safe to run before it is fast to build. Syltech works across the full lifecycle so teams ship intelligent software without trading away reliability, and hands back platforms they can actually operate.

security · policy-guarded tools

// Tools an agent may call, gated by policy.
platform.Register("db.query", queryHandler,
    policy.Require("resource:read"),   // Ory Keto check, per-tenant
    policy.RateLimit(100, time.Minute),
)

agent · governed loop

// Autonomy that stays safe to run in production.
for step := range agent.Plan(ctx, goal) {
    if err := budget.Guard(step); err != nil {
        return err          // hard spend + time ceiling
    }
    out := durable.Do(ctx, step)  // crash-safe, exactly-once
    agent.Observe(out)            // grounded + evaluated
}

Stack

What Syltech builds with

Languages

GoTypeScriptPythonRust

Cloud

KubernetesOpenStackAWSAzureDockerFirebase

Platform / IaC

DaggerDBOSTerraformOpenTofugophercloud

Security & IAM

OryRBACReBACOIDC / SSOTrivyOpenSCAPVault

Data & Observability

PostgreSQLRedisMongoDBOpenTelemetry

AI

LangGraphLangChainMCPClaude / OpenAIRAG

Engagements

How Syltech works

Advisory

Architecture & strategy

Architecture reviews, platform and security strategy, and hands-on guidance — you leave with a clear plan and the risky parts named before they cost you.

Delivery

Scoped project builds

Golden-image pipelines, cloud control planes, authorization layers, and AI integrations — shipped, documented, and handed back production-ready, not a prototype.

Fractional

Embedded platform team

Ongoing platform and infrastructure engineering, embedded alongside your engineers — raising the floor on reliability and delivery, and keeping it there.

Field notes

Essays and guides

Medium

RBAC Is Technical Debt: Zanzibar-Style Authorization in TypeScript

Medium

Production Authorization: The Dual-Write Hazard, and a Blueprint for Machine Identity

Substack

What Happens When Your Monitoring Can't See Behind the Firewall

Substack

Building Sentry: A Container Security Scanner with Dagger

Open source

Public, production-grade code

WatchDog · dashboard

leash · capping a run

$ leash --max-cost 0.10 -- ./agent.sh
step 1   $0.03   ok
step 2   $0.07   ok
step 3   $0.10   budget reached
halted ./agent.sh at $0.10 / $0.10 cap

rerun · durable step

// Runs once. Survives a crash and resumes.
val, err := rerun.Do(w, "charge",
    func(ctx context.Context) (string, error) {
        return chargeOnce(ctx, order)
    })

Designed, shipped, and running in production.

leash

Durable spend governor for AI agents; caps cost and survives restarts.Go

rerun

Durable execution for Go: workflows that survive a crash and resume.Go

drover

Durable, budgeted agent runner: plan, act, observe, survive.Go

WatchDog

Open-source infrastructure monitoring.Go

RustSensei

Offline Android tutor with on-device LLM inference.Kotlin

ctx-forge

Deterministic context CLI for AI coding agents.Rust

All repositories on GitHub →

Founder

Sylvester Francis

Principal Software Engineer and Architect with 8+ years building cloud platforms, golden-image and authorization systems, and applied AI. Builds secure, reliable systems, and now runs Syltech.

FocusPlatform engineering · DevSecOps · Identity & authorization · Applied AI

LinkedIn → · Medium → · Substack → · GitHub → · Hugging Face →

When to call

Bring in Syltech when…

Identity

Authorization is sprawling

RBAC copy-pasted across services, multi-tenant permissions no one fully trusts, and machine identity bolted on late.

Delivery

Shipping is slow and risky

Ticket-based provisioning, no golden-image pipeline, and releases the team crosses its fingers on.

Agents

Agents are loose in production

LLM agents running without spend limits, durability, or any record of what they did.

If any of that sounds familiar, book a call →

Contact

Open for work

Have something to build?

Syltech takes on new engagements and consulting across AI, platform engineering, and full-stack delivery. Write with a sentence about the problem, and you will hear back.

Book a call →

Or book a paid 90-minute architecture review →

Prefer email? [email protected] · LinkedIn →

Syltech uses what you send only to reply, and never shares or sells it. Privacy →

Platforms, infrastructure, and the AI that runs on them.