Benoit Meunier — Senior Software Engineer, Montreal

Selected professional work

26 years experience · Montreal

Bminty · 2020 — Present

LNG Platform

Sole author of two greenfield enterprise SPAs, owning every layer from API contract to rendered interface: 75+ entity pages, 50+ advanced data grids and 50+ multi-tab forms with validation, dirty tracking and optimistic concurrency, over a custom state architecture of 25+ composable React Context providers and no external state library. A 100+ component design system with a hot-swappable Dark / Light / Branded theme, GraphQL subscriptions over WebSockets, fragments generated from the grid fields actually on screen, field-level permissions across 60+ business entities, and 50+ locale files across 6 regions.

React 19TypeScriptGraphQLApollo Client 4ViteTailwindCSS 4Radix / Shadcn

An AI engineering workspace: a developer at a multi-monitor desk surrounded by panels describing open-source agent and developer tooling projects

Bminty · 2020 — Present

AI developer tooling

Designed and built ai-api, a Go service that turns free-form, multi-locale, tenant-specific questions into the typed, validated filter arrays the platform's search subgraph consumes — layered configuration, pluggable LLM and vocabulary backends, and a full test pyramid including parity tests against the live search API. Alongside it, an MCP server that exposes the frontend's component, hook and schema catalog to GitHub Copilot, plus the Copilot customization layer — agents, skills and instructions — that every engineer on the project works through.

GoNodeTypeScriptModel Context ProtocolGitHub Copilot

Interactive television interfaces across a smart TV and three handheld screens, showing an on-now banner, a continue-watching row and app and channel navigation

LVL Studio · 2007 — 2020

Interactive TV

12.5 years owning TV user experience end to end — set-top box, mobile and desktop — on Microsoft Mediaroom and Bluestreak, at a Montreal studio specialised in interactive television. Shipping minimal code into resource-constrained embedded set-top hardware, where memory and render budgets are fixed and every release passes the carrier's own deployment gate. Carrier-grade delivery to millions of subscribers at Bell Canada and Telus, plus Time Warner Cable and the Hayu streaming service.

Microsoft MediaroomBluestreakSet-top boxiOSAndroidWeb

Open source & personal projects

4 repositories · all public

Noctis

An autonomous research system for trading strategies. An LLM agent drives a formulate / match / optimize / decide cycle through a curated tool registry, writing one reviewable Python file per candidate and validating it in a subprocess before it can land. Validation is enforced in code rather than by convention: walk-forward splits, out-of-sample on the temporal and symbol axes, overfit-gap guards and a tested no-lookahead invariant. ~140 typed modules, 93 test modules, 93% coverage. Paper-only by construction, behind two independent safety gates.

93%

Coverage

140

Typed modules

Test modules

PythonPydanticpandasNautilusTraderDataBento

lexiqr

A deterministic Python library for B2B SaaS teams whose tenants each speak their own private language. A tenant's lexicon maps its jargon onto canonical database entities, and free-form prompts come back as entity-identified objects: exact and typo-tolerant matches with character spans and scores. Chosen over embeddings and NER for explainability and a CI-enforced p95 under 10 ms, with determinism verified across operating systems and Python versions against a committed golden report, and 45 test modules against 15 source modules.

p95 < 10 msPythonRapidFuzzmypy strict

noctis-ollama

Turns a bare GPU box into a keyless local LLM backend the Noctis research engine can drive tools through, at no cost per token. One command installs Ollama, provisions the models a manifest declares, and proves each one against a 7-gate contract test asserting exactly what the consuming code parses: tool-call round trip, OpenAI /v1 parity, 100% GPU residency. Fits a 14B model with a pinned 16k context into 12 GB of VRAM via q8_0 KV-cache quantization and flash attention.

ShellPythonOllama

beautify-md

A Claude Code skill that measures, audits, rewrites and authors markdown — READMEs especially — so humans enjoy reading it. Six layers, from why the document exists down to what GitHub actually renders, every rule traced to readability research, and a bundled bash and awk script that keeps the numbers honest. It needs nothing but bash and awk: no build step, no keys, no external services.

ShellawkClaude Code

About

Montreal · French & English

My personal projects are where I get to build the thing I actually want to build — no scope negotiation, no compromise. Lately that means AI: LLMs, agents, and systems that run on their own.

Noctis is the clearest example. It's an autonomous investment research system on a daily cycle — overnight it researches and generates trading strategies, during market hours it paper-trades them to see what holds up, and every trade gets recorded so the results feed the next night's research. The loop is the point. Strategies get proposed, tested against real conditions, and revised without me in the middle of it. It runs on noctis-ollama, which turns a bare GPU box into a local model backend it can drive tools through — so a research cycle that thinks all night costs nothing per token.

Professionally, I'm a Senior Software Engineer at Bminty.

Career

2020 — Present Senior Software EngineerBminty (formerly ProConsultant Informatique) · Montreal
2007 — 2020 Senior Software EngineerLVL Studio · Montreal
2001 — 2007 Software EngineerPurelogic · Montreal
1999 — 2001 Web DeveloperSid Lee · Montreal

ben