AI × finance engineering student. At Hubelia, I designed and shipped agents that work every day. This site is the living proof.
Projects
Agentic video studio
20+ AI agents that illustrate a short video, automatically.
Make short videos (Reels / TikTok) educational and premium-looking without an editor: the illustration has to explain what is being said, not repeat the subtitles.
- End-to-end pipeline: local transcription → detection of free zones (never covering a face) → storyboard → HTML/GSAP scene rendering → frosted-glass composite on the video.
- Orchestration of 20+ specialized Claude sub-agents, with judge ↔ fixer loops that mix LLM decisions and deterministic measurements.
- Deterministic, frame-accurate rendering (seeked GSAP timeline, ProRes alpha clips).
Working end-to-end pipeline (advanced POC), with real videos produced.
Agentic internship hunter
Six agents source the offers, one CV and one cover letter per offer, an automated and verified submission.
Hundreds of relevant offers scattered across a dozen recruiting portals, each with its own form, accounts and attachments: impossible to keep up by hand without missing something.
- Six sourcing sub-agents (job boards, banking and finance, industry, large groups, AI startups, Montréal) that return verified offers, scored out of 5 and deduplicated against everything already sent.
- For each selected offer, a CV and a cover letter tailored to the position, generated then proofread, as PDFs ready to submit.
- Automated submission with Playwright on most portals (Workday, Taleo, Talentsoft, SmartRecruiters, Welcome to the Jungle…): accounts, forms, attachments, with file checks before submitting and a report for every submission.
297 applications prepared as of October 25, 2026, in successive waves since June; first interviews obtained. The same tooling keeps my LinkedIn profile up to date.
Demo video studio
An app URL → a 1080p demo video, rendered frame by frame.
Produce demo and presentation videos of a web product without filming or editing, from the application's real data (figures read from the DOM, never made up).
- Capturing the app with Playwright (HD screenshots, DOM extraction, brand accent detection, cold-start handling).
- A pipeline of ~10 agents (scout → designer → animator → judge ↔ fixer) composing a JSON contract, in 3 formats (demo, guided tour, story).
- 100% deterministic rendering engine (each frame depends only on its index) → headless Chromium → ffmpeg, with guardrails (JSON / Lottie validation) and tests.
Operational: 5 videos actually rendered, ~25 green tests, "story" format validated.
Generative merch
Design + product mockup generated without a graphic designer.
Produce a unique merch design per brand or slogan at near-zero cost, instead of paying a graphic designer for every variation.
- Agentic worker ↔ adversarial verifier pipeline, from the brief to the print-ready visual and the product mockup.
- Copyright firewall: the brand is never sent to the image generator, only a neutral visual brief is.
- Tight cost control (caching, model choice) — unit cost measured and optimized.
Measured cost ~$0.40/design (cache hit), multi-model pipeline working in staging.
E-commerce storefront
The shop that sells the generated products — one design on every kind of item.
Give the AI-generated product line an e-commerce storefront: a modern shop where each design comes on several items.
- Customization of a modern e-commerce base (Medusa v2 / Next.js 15 / React 19): dark theme, brand identity, French content.
- A storefront built to sell the products from the "Generative merch" pipeline, on t-shirts, totes, bottles, cushions…
- Complete e-commerce journey provided by the base (product → cart → Stripe checkout).
Storefront prototype; examples of products for sale in the gallery below.
From the meeting to the CRM
Meeting audio → bilingual summaries → CRM record.
Capture every sales meeting without retyping: a clean report for the team and a presentable one for the client, attached directly to the prospect's record.
- n8n chain of ~57 nodes: transcription + diarization → structured LLM generation → write to the CRM.
- Business vocabulary injection (contacts, employees) to make name recognition reliable.
- Bilingual (FR/EN) and multi-format summaries (internal / client), triggered from the CRM.
In internal use; multi-language and multi-type summaries operational.
CRM enrichment
B2B decision-makers enriched automatically, without wasting credits.
Replace the manual search for B2B decision-makers with automatic CRM enrichment, while keeping paid data credits under control.
- Free lookup of a company's contacts, then paid enrichment only for the new ones (0 credits if nothing is new).
- Asynchronous retrieval of phone numbers (callback), with automatic refresh of the record.
- UX integrated into the CRM (decision-maker selection pop-up, badges).
3 workflows in production; ~370 companies processed.
Email triage assistant
Two inboxes sorted by Claude, a Telegram digest every morning, drafts on request.
My personal project: never miss an important email (school, applications, interviews) drowned in the noise of two inboxes — without handing my mail to a third-party service.
- Automatic hourly triage: deterministic rules first, then classification by Claude; when in doubt, a question with buttons on Telegram rather than a guess.
- A concise digest every morning (scheduled PC wake-up, automatic catch-up if the wake-up fails); from the phone, a typed instruction becomes a reply draft in the right thread.
- In-house MCP server: any Claude session can search, read and prepare a draft — with no sending or deleting code at all, and email content fenced against prompt injection.
In daily use on my two inboxes; built with TDD (250+ automated tests).
Monte Carlo pricing
From the specification to the quote price, through 10,000 simulations.
Price software and AI quotes with a controlled margin despite the uncertainty on hours, change requests and LLM usage costs.
- Accounting foundation in the ERP (accounts, cost centers, control server script).
- Specifications generated by an LLM from meeting reports.
- Monte Carlo engine (10,000 draws) calibrated on real variability → recommended price (Q95) + discount grid, delivered in a dashboard.
Chain validated end to end, delivered as a turnkey kit.
ERP operations agent
An agent that reads and updates projects & tasks in natural language, safely.
Give conversational access to the ERP's delivery data (projects, tasks, todos) and surface slippages, without going through a heavy interface.
- Read client optimized for LLM context (projection of high-signal fields, aggregation in code).
- Write layer guarded by a resolve → preview → confirm → write protocol (dry run, reversible vs destructive distinction).
- Agent identity and rules pinned down (request → document type routing, tiered retrieval).
A tool that genuinely operates; "read-only by design" architecture (the most mature on the engineering side).
DataInvst
Web platform for portfolio management, with real-time market data.
My personal project (2025): a full-stack platform to track and analyze an investment portfolio, from stocks to crypto — my AI × finance playground.
- ASP.NET Core 8 (C#) back end + Entity Framework Core / SQL Server: portfolios, positions, transactions, price history.
- Finnhub API integration for real-time quotes (stocks & crypto) and market history.
- Documented REST API (Swagger), secure authentication (BCrypt hashing), admin area.
A working full-stack application — my most accomplished personal project.