Résumé

Seeking Summer 2027 internships · Backend · Fintech · ML · TypeScript

Lukas Panos

Computing @ Queen's University. I build ML models

Third year · Software Design specialization · Kingston, ON

See the work Download résumé GitHub
01 About
Portrait of Lukas Panos
// lukas.panos — kingston, on

I write software that has to hold up: under retries, under crashes, under out-of-sample data, and in front of real users.

I'm a third-year Computing student at Queen's University specializing in Software Design, with a strong interest in backend engineering and fintech. I came to software from the business side: working in real estate marketing, I kept building tools for agencies, and that turned into Agency Copilot.

Payments in particular is where my business experience meets the challenge of building reliable systems. I'm also exploring machine learning as an ML Engineer with QMIND, and I have a personal interest in markets and trading. I'm fluent in English and French.

I like problems where the obvious solution quietly breaks. My work spans payment ledgers with invariants enforced in the database, RAG pipelines that cite their sources, and ML experiments that report the inconvenient result.

Seeking Summer 2027 SWE internships in
  • Backend
  • Fintech & Payments
  • Machine Learning
  • JavaScript / TypeScript
  • Full-stack
0tests against a real Postgres in my ledger service
0federated ROC AUC, beating every hospital trained alone
0faster client reporting with Agency Copilot (2+ hours down to under 5 minutes)
0real estate client accounts managed at Versatile Agency
02 Selected work

Things I've built.

01 Active users

Agency Copilot

SaaS platform for marketing agencies

Problem
Agencies run ads, CRM pipelines and team tracking across 3+ disconnected tools and manual spreadsheets, and client reporting eats hours per account.
Built
A multi-tenant command center that unifies Meta Ads data, GoHighLevel CRM pipelines and team management. Its AI reporting engine, built on the Claude API, writes executive briefs and client-facing PDF reports, cutting reporting from 2+ hours to under 5 minutes per account.
  • Next.js 16
  • React 19
  • Supabase
  • Clerk
  • Puppeteer
  • Recharts
  • Claude API
  • Vercel
Live site ↗
02

DocMind

AI document intelligence (RAG)

Problem
Long PDFs are hard to search, and most chatbots answer without showing where the answer came from.
Built
A full-stack RAG app. It splits uploads into ~500-token overlapping chunks, embeds them with OpenAI and stores them in a pgvector store built from scratch on Supabase with an ivfflat cosine index. Claude streams answers that cite the exact chunks they came from, and each document keeps its own conversation history.
  • Next.js 16
  • React 19
  • Supabase pgvector
  • OpenAI embeddings
  • Claude
  • Tailwind v4
Live demo ↗ Source ↗
03

Double-Entry Ledger

Payments infrastructure

Problem
Ledgers rarely fail on the happy path. They fail under concurrent retries, crashes mid-transaction and silent balance drift.
Built
A payments-grade API. Entries are append-only and balances are derived from them, never stored. Postgres enforces a zero-sum rule per currency with deferred constraint triggers. It also has atomic idempotency keys, two-phase authorization holds with partial capture, and hash-chained transactions. Proof: 218 tests, and a chaos runner that killed the server 15 times across 30,179 operations without breaking an invariant.
  • Python 3.12
  • FastAPI
  • PostgreSQL 16
  • Docker
  • pytest
  • Hypothesis
Source ↗
04

Volatility Regime Classifier

ML forecasting & strategy validation

Problem
Can a model forecast Nasdaq-100 volatility regimes 10 days ahead? And if it can, is that forecast actually worth money?
Built
A walk-forward pipeline with leak-free labels. It reached 71.4% out-of-sample accuracy against a 67.9% persistence baseline over 3,200 trading days. The strategy built on it cut volatility from 21.4% to 8.8% and max drawdown from 35.3% to 13.7%. A Monte Carlo layer then showed the timing edge was indistinguishable from chance (permutation p = 0.46), and I reported that result rather than stopping at the good-looking one.
  • Python
  • scikit-learn
  • XGBoost
  • pandas
  • NumPy
  • Matplotlib
Source ↗
05QMIND × Distributive

FedHeart

Federated learning across hospitals

Problem
Hospitals can't pool patient data, but each one alone has too little data for a strong model.
Built
A dependency-free federated training package that runs locally or on Distributive's DCP. Four hospitals train a shared heart-disease model and never send raw data. It survives a hospital dropping out mid-training and supports optional secure aggregation, differential privacy and per-hospital personalization. The federated model reached 0.886 AUC, beating every hospital alone and matching the pooled-data upper bound.
  • JavaScript
  • Node.js
  • DCP
  • FedAvg
  • Differential privacy
  • Secure aggregation
Source ↗
06

Melo AI

Music recommendation engine

Problem
Pure audio-similarity recommenders cross genre lines. A pop song would surface a tempo-matched classical piano piece.
Built
A recommender that first filters to songs in the same genre family, then ranks them by cosine KNN over 11 Spotify audio features across ~268k tracks. It has popularity-ranked typeahead search, 30-second iTunes previews and one-click links to Spotify.
  • Next.js 14
  • TypeScript
  • Tailwind
  • FastAPI
  • scikit-learn
  • pandas
Live demo ↗ Source ↗
03 Experience

Where I've shipped.

Oct 2026 — Present

Quantitative Analyst @ QUANTT

Kingston, ON

  • Project team member building an algorithmic futures trading bot, translating discretionary trading concepts into systematic, rule-based signals that can be backtested and automated.
  • Code the bot's trading logic in Pine Script on TradingView, including entry and exit rules, position sizing and strategy backtests. I also contribute to strategy research and performance evaluation.
  • Run the strategy on prop-firm funded capital, where a fixed evaluation fee caps the downside. Risk management is built around each firm's drawdown and daily-loss rules.
Sep 2026 — Present

Machine Learning Engineer @ QMIND

Kingston, ON

  • Selected for a design team partnering with Distributive, a Kingston distributed-computing company, to build a proof-of-concept federated learning package on its DCP platform for predicting patient response to biologic treatments.
  • Co-built a full-stack federated learning system at the QMIND hackathon. It trains a heart disease model on patient data from multiple countries without pooling it, using differential privacy and secure aggregation.
Mar 2024 — Oct 2025

Senior Marketing Specialist @ Versatile Agency

Montreal, QC

  • Managed 30+ real estate client accounts across Meta, Google and YouTube, overseeing $40K+ in monthly ad spend with full accountability for CPL and ROAS.
  • Rebuilt the agency's operational infrastructure, including GHL pipelines, ISA workflows, lead-routing automations and reporting SOPs. This cut manual reporting time by 5+ hours per week.
Mar 2023 — Jun 2025

Founder & CEO @ OXP Media

Vancouver, BC

  • Founded and ran a real estate marketing agency solo, reaching $10K monthly revenue and $7K+/month in managed ad spend. I handled all sales and client acquisition from zero and kept 70%+ retention.
Expected Apr 2028

B.Sc. Computing, Software Design @ Queen's University

Kingston, ON · GPA 3.61 / 4.0

04 Toolkit

Skills

Languages

TypeScript · JavaScript · Python · SQL · HTML/CSS

Frameworks

Next.js · React · FastAPI · Tailwind CSS · Pandas · NumPy

ML & AI

scikit-learn · XGBoost · Keras · fast.ai · time-series CV · Claude API · OpenAI embeddings · RAG · pgvector

Tools

PostgreSQL · Supabase · Docker · Git · Vercel · Clerk · Puppeteer · Pine Script / TradingView

Languages (human)

English · French (fluent in both)

05 Contact

Let's build something that holds up.

I'm seeking software engineering internships for Summer 2027 in backend, fintech, machine learning and JavaScript/TypeScript. If you're hiring, or just want to talk systems, markets or ML, I'm always open to connecting.