Three years in production, from backend architecture to continuous delivery. NestJS and Node.js on the server, React and Next.js on the client, PostgreSQL underneath. API security and secrets handling applied from the design phase, not bolted on after. Antananarivo, Madagascar.

Three projects follow. Each had one thing it had to get right, and that thing is what its section shows.

A movie's score is the attention it has received, each interaction discounted by how long ago it happened:

$$S(t) = \sum_i w_i, e^{-\lambda (t - t_i)}, \qquad \lambda = \frac{\ln 2}{t_{1/2}}$$

Factor the constant out and what is left is an accumulator that only ever needs an O(1) update — no rescan of history, ever. And since $e^{-\lambda t}$ is positive and shared by every movie, it cannot change their order: sorting by the accumulator is enough. That is why nothing has to run on a schedule. The ranking decays correctly by construction.

The accumulator overflows a float64 quickly, so it is stored as $\log A$ and updated with log-sum-exp, which keeps it finite for decades.

Three half-lives run in parallel, giving three views of the same journal:

Next.js 16 · Postgres · Drizzle  ·  repo  ·  live

The textbook case, "I can see what I own", proves nothing. The one worth building is delegation: an agent holds a mandate on a listing they do not own, and it lapses on its own. Every cell below is one named test.

The suite applies the production schema verbatim inside an in-process PostgreSQL. No container, no external service, no server to start.

Three traps it walks into on purpose: policy recursion between two tables, which needs a SECURITY DEFINER function to break; the discipline that function then owes, since it is an RLS bypass; and the fact that GRANT is not RLS — a mandated agent may edit a listing but must not reassign it, which WITH CHECK cannot express because it only ever sees the row as it will be, never as it was.

SvelteKit · Supabase · PGlite  ·  repo  ·  live

A format takes its entrants and returns the matches to play plus the edges saying where each winner and each loser goes next. Nothing else.

graph LR
  M1[Match 1] --> SF[Semi final]
  M2[Match 2] --> SF
  M1 -.-> LB[Losers bracket]
  M2 -.-> LB
  SF --> F[Final]
  LB --> F
  classDef signal fill:#FBBF24,stroke:#B45309,stroke-width:2px,color:#0D1117;
  class F signal;

Solid edges carry the winner, dashed ones the loser. Everything downstream reads that graph and knows nothing about brackets: propagating a result, cascading a correction through every match it invalidates, computing live title odds, deciding the tournament is over. Adding a format is one file and one registry entry.

The odds are exact on edge formats — one dynamic-programming pass over the DAG, each pairing weighted by the Elo expected score, no sampling error. Round robin has no edges, so outcomes are enumerated exactly while few enough remain and sampled beyond that. Both are covered by property tests asserting the probabilities sum to one and stay monotone in rating.

NestJS · React · SQLite  ·  249 tests  ·  repo  ·  no public demo — it runs from one Docker command

Co-founder and full-stack developer — NEXTRI Antananarivo · May 2026 – present

A collective of three software engineers, hired as a team or as a single reinforcement. Its technical lead. We designed an internal application from scratch, the three of us — NestJS/TypeScript API, PostgreSQL schema, CI/CD pipeline, licensing — from design through tests, with REST endpoints secured and secrets handled from the design phase. It is an internal tool: not deployed, no users.

Full-stack application developer — Stellar-IX (Axian Group) Antananarivo · September 2022 – January 2025

Cut a VMware monitoring application's average data load from 4s to 300–500ms by rewriting its backend in FastAPI — vSphere REST for routine operations, pyVmomi where the SOAP SDK was required — with caching and ClickHouse query optimisation.

Shipped more than ten full-stack modules to production with Docker CI/CD and systematic code review, after getting TypeScript accepted for every new module on a codebase that ran on legacy ejs and Node. That argument took several team discussions, and it is the reason the title above says TypeScript. Contributed to the team's technical decisions — architecture, data patterns, code quality — in a ten-person multidisciplinary team working Agile.

Mobile developer, internship — Ingenosya Madagascar Antananarivo · August 2021 – February 2022

Built the entire backoffice of an audio-visual e-learning platform, from scratch, for a European client, in a team of eight. Followed it from design to delivery: client meetings, then successive releases to staging, pre-production and production.

Master MIAGE — Mobiquity, Databases and Systems Integration (MBDS). Université Côte d'Azur & IT University, Antananarivo · 2022 – 2025. Bachelor in computer science, Web and Design. IT University, Antananarivo · 2018 – 2022.

DevSecOps Professional — Practical DevSecOps, January 2024. A three-day instructor-led course with no examination, so: a course, not a certification.

Main — TypeScript · React · NestJS · PostgreSQL Backend beyond it — Node.js with Express · Java with Spring Boot · Python with FastAPI and Flask Frontend — React · Next.js · Angular Data — PostgreSQL · MySQL · MongoDB · ClickHouse API and security — REST design · endpoint hardening · secrets handling · DevSecOps DevOps — Docker · CI/CD · Git Architecture and quality — system design · clean code · refactoring · code review · tests AI-assisted development — Claude Code, for code, tests and documentation

Both demos are probed every six hours from GitHub's runners, five requests each, and the figure above is the median — a measurement, not a claim. A failed probe breaks the line rather than being drawn as a fast response. The workflow rewrites a single-commit metrics branch each time, so no machine noise ever reaches this history.

LinkedIn  ·  rasatasitraka2@gmail.com

Malagasy (native) · French (fluent) · English (intermediate to professional, written)