Loading...

Clarity

Dennis Sirois

New Durham, NH

I help founders, engineering leaders, and operators put Claude and other LLMs to real work — building custom MCP servers, multi-agent systems, and autonomous workflows that survive contact with production. Not prompt-engineering content; the deeper layer where most teams get stuck.

I've authored three custom MCP servers (Node + Python, across stdio and ZeroMQ transports), shipped a 3-layer multi-agent architecture with one Claude session supervising another via file-based escalation, and built a hook-based guardrail system that intercepts user prompts to inject urgent-state context. My largest MCP server exposes 81 tools to Claude over the Chrome DevTools Protocol. My autonomous trading bridge gates every order-placing tool behind an environment-variable kill switch and writes a JSONL audit trail of every call.

The day job is AI Systems Engineer at a managed IT firm serving municipalities across New England — building MCP-based integrations for helpdesk, compliance, and workflow systems, managing their Claude Teams account, and shipping automations for government-adjacent IT operations. Previously Senior Systems Administrator at a publicly traded utility for six years. Twenty-five years total across utility, FDA-regulated, SSAE-16/SOC, and DCAA/DCMA-compliant environments. That background is why the AI work is engineered the way it is: structured tool schemas, safety interlocks, audit logging, versioned modules, supervision layers — not one-off prompts.

Calls I'm best at:
You're building an MCP server (or thinking about it) and want a second set of eyes on tool design, transport choice, or schema discipline before you ship it.
Your Claude Code setup is a pile of slash commands — you want to graduate to skills, subagents, hooks, and custom MCPs without breaking what works.
You're standing up an autonomous or semi-autonomous agent loop and need to design the safety surface (kill switches, gating, audit trails, supervision) before it's loose on real systems.
You're integrating Claude with a desktop app, trading platform, or proprietary system via computer-use, CDP, or a custom bridge and the official docs run out before your problem does.
You're a CTO/founder evaluating "build vs. buy" on agentic tooling and want a practitioner's read on what's actually achievable in 30 days vs. what's a 6-month research project.
You're an IT leader being asked to deploy Claude/AI agents inside a regulated environment (SOC 2, SSAE, FDA, utility/OT) and need a sober conversation about what that actually looks like.
I work better in 30-minute calls than 30-page proposals. Book a call — first 5 minutes are on me if it's not a fit.

$2.92/min

—
0 Calls

—

0 Reviews

—

Avg. call

Information Technology Artificial Intelligence

Call me to talk about

Claude Code, MCP & Custom Agentic Tooling

I've authored three production MCP servers across Node and Python, using stdio and ZeroMQ transports. My largest exposes 81 tools to Claude over the Chrome DevTools Protocol. I've also built custom Claude skills, subagents, slash commands, and a UserPromptSubmit hook that injects supervisor-agent context into the running session.

Book a call if:
You're scoping or building an MCP server and want to pressure-test tool boundaries, schema design, transport choice, and skill/subagent layering before you commit.
Your Claude Code project has outgrown ad-hoc prompts and you want a real architecture: project-level CLAUDE.md, scoped skills, subagent-as-tool, hook-based guardrails, file-based memory.
You're trying to wire Claude to a proprietary or desktop app (CDP, ZMQ, custom bridge) and the standard examples stop where your problem starts.
You're evaluating Anthropic Agent SDK vs. Claude Code vs. raw API for an internal tool and want an honest comparison from someone who's shipped in all three.
Outcome: a concrete next-step decision — schema, transport, scope, or staging — written down before the call ends. Not a sales pitch for more consulting.

Multi-Agent Systems & Autonomous Loops (Safely)

I've built a three-layer multi-agent system with a development layer, an execution agent running an autonomous loop, and a supervision layer where one Claude session watches another via file-based escalation and an injected URGENT.md hook. My autonomous trading bridge runs Claude on a self-pacing tick, with hard guardrails baked into the C# arbiter and Python bridge.

Book a call if:
You're about to point an agent at production and want to design the safety surface first: env-var kill switches, mutating-vs-read tool separation, gated arming, monotonic state guards.
You need an audit trail for an autonomous agent (we should talk about what to log, where, and how to make it survive a compliance review).
You're designing inter-agent communication — file contracts, supervisor → executor channels, escalation patterns — and want to skip the obvious failure modes.
You're stuck on an agent loop that "almost works" but drifts, loops, or hallucinates state, and you want a forensic look at the run.
Outcome: a sketched safety architecture, named failure modes, and the 2–3 things to instrument before you run the loop again.

AI in Regulated & Critical-Infrastructure Environments

I'm currently building MCP-based AI integrations for helpdesk, compliance, and workflow systems at a managed IT firm serving municipalities across New England — government-adjacent operations where controls and accountability aren't optional. Previously six years as Senior Systems Administrator at a publicly traded utility. I've designed DR plans that passed SSAE-16 audit, maintained 99.99% uptime for an FDA-regulated client under SLA, and held the line on the boring controls (AD, PKI, backups that actually restore) that make any AI deployment defensible.

The interesting question for most regulated buyers isn't "should we use AI?" — it's "what can we ship without ending up on the wrong end of an audit?"

Book a call if:
You're an IT leader at a utility, hospital, financial firm, or regulated SMB being pushed to deploy LLMs and you need to scope it without blowing the controls you already have.
You're drafting an internal AI policy and want a practitioner's read on what's actually enforceable vs. theater.
You're evaluating a "Claude inside our environment" vendor pitch and want a quiet second opinion on data exposure, network architecture, and audit posture.
You need to bridge an IT/OT divide where the OT side will not — under any circumstances — touch the public internet, and you want to talk through air-gapped, local-LLM, and proxy patterns.
Outcome: a realistic, scoped path forward — what's safe to pilot in 90 days, what to defer, and the 3 controls you should not skip.

AI for Trading, Quant Workflows & Desktop Automation

I've wired Claude directly into NinjaTrader 8 over a custom ZeroMQ bridge — placing, modifying, and cancelling orders with gated arming, session loss limits, and per-call audit logs. This runs on a sim account with one contract; the engineering is the point, not the P&L. Separately, I've trained Soft Actor-Critic (SAC) agents in Gymnasium against the same bridge using Stable-Baselines3, with walk-forward validation, parallel hyperparameter sweeps, and live model inference driving the platform.

I'm not selling a strategy. I'll talk shop on the engineering and risk surface around AI in trading workflows — where it earns its keep, where it absolutely doesn't.

Book a call if:
You're a prop trader, quant, or fintech founder thinking about LLM-driven trading or research workflows and want a clear-eyed read on what's real vs. what's marketing.
You're integrating Claude with a desktop trading or charting app (NinjaTrader, TradingView, MT5, Sierra) and need a working pattern for the bridge.
You're scoping an RL pipeline (SAC, PPO, custom env design, walk-forward) for a small team and want feedback on the architecture and reward design before you spend a month on it.
You want to discuss safety design — kill switches, sandbox vs. live, audit trails — for an AI agent that has order-entry authority.
Outcome: a sharper picture of where AI helps in your workflow, where it doesn't, and what the next 30 days of work should actually look like.

AI Session Continuity & Institutional Memory

AI sessions are ephemeral by default — every decision, rationale, and lesson learned in a working session evaporates when the context resets. I've built and use a structured end-of-session workflow (a custom Claude skill) that runs a disciplined sweep before any session closes: inventorying what changed, filing decisions into the right permanent documents, hunting down stale explanations elsewhere in the project and updating them, and producing a clear handoff summary before sign-off.

The standout piece is what I call the rationale blast radius — when a decision changes, the sweep actively finds every other document that references the old reasoning and updates it. Old documentation that quietly contradicts current practice is one of the most expensive problems in AI-assisted teams, and most people don't realize it's happening until it costs them.

Book a call if:
Your team uses Claude Code or another AI assistant for real work and keeps losing context between sessions — decisions get re-litigated, constraints get forgotten, work gets repeated.
You're deploying AI tools internally and want a framework for maintaining institutional knowledge across session resets and team handoffs.
You're building a multi-session or multi-agent workflow and need a durable memory and continuity design before you scale it.
You've experienced the "the AI forgot why we made that decision" problem and want a systematic fix, not a workaround.
Outcome: a session continuity design — what to capture, where to store it, how to keep it consistent — that fits your existing project structure and doesn't require rebuilding everything.

Answers 0
Reviews 0

Answers

Reviews

Member since June 2026