AI & Data Science Engineer · ESI-SBA, Algeria
I'm Akram Rahali, a fifth-year Artificial Intelligence & Data Science engineering student. I take models (graph neural networks, vision transformers, detectors, retrieval pipelines) all the way to the product someone actually opens.
Graph, vision and language models, often straight from a paper, trained and evaluated in PyTorch.
PyTorch · PyG · DINOv2 · YOLOv8 · CTransPath
Inference behind typed APIs, results persisted, everything containerised and reproducible.
FastAPI · NestJS · PostgreSQL · ChromaDB · Docker
Maps, 3D viewers and mobile screens that turn predictions into decisions.
React · Three.js · Jetpack Compose · React Native
Four end-to-end builds. Each one is shown in its own colors, the palette the app actually ships with.
Dairy sustainability forecasting for Ireland's 26 counties.
A direct implementation of a spatio-temporal GNN paper: herd records become a county graph, the model forecasts each county's sustainability score to 2036, and policy levers run counterfactual what-if simulations.
Read the paper · arXiv:2512.19970Pillars: reproductive efficiency, genetic management, herd health and herd management. Forecasts and simulations are persisted in PostgreSQL and served to an interactive atlas with light and dark themes.
Formation flight that emerges from local, learned decisions.
Swarm episodes simulated in PyFlyt are turned into communication graphs. Graph neural networks learn formation assignment and each drone's local setpoint, so control can move from one central planner to every agent, and both modes are replayed side by side.
Each PyFlyt timestep becomes a graph: drones as nodes, communication links as edges.
Formation assignment (who goes where) and local setpoint prediction (how to get there).
A Three.js viewer served by FastAPI runs dual simulations and scrubs through saved runs.
Stacked-timber volume from one photo, fully on device.
Point the phone at a log pile: a YOLOv8 segmentation model finds every log end, two reference points fix the real-world scale, and the app returns per-log diameter and total stack volume. No server, no signal needed in the forest.
YOLOv8 log-end segmentation, tiled 1 / 4 / 9 / 16 so dense piles keep small logs.
Two draggable reference points and a known distance set pixels per centimetre.
Diameters × log length give per-log and total volume, with original, mask and YOLO views.
Botanical relevés from a single field photo.
Ecologists normally estimate plant cover by eye. Here a photo of a vegetation plot is embedded with a fine-tuned DINOv2 vision transformer, and lightweight heads score species abundance, dominance and Braun-Blanquet cover classes.
Medical imaging, agentic retrieval and a full livestock platform, built with the same model-to-product approach.
Slide-level tissue classification. Patches are embedded with CTransPath, classified by classical and semi-supervised models, and their votes aggregated into one slide label, all in a Gradio app.
Patches → CTransPath → SSL + boosting → slide vote
An agentic reader for scientific literature: it searches arXiv, ingests papers into a local vector store, and answers from session-scoped retrieval refined by a cross-encoder reranker.
arXiv search → ChromaDB → rerank → grounded answer
Livestock operations for Algerian farms. Farmers, veterinarians and zootechnicians share one app for herd records, vet visits, feed rations with nutritional totals, and professional contracts.
Role-based mobile app → NestJS API → PostgreSQL
DOT Khenchela
Designed and built a full-stack telecom billing platform on a local client–server architecture: residential and enterprise subscribers, bimonthly consumption and fee calculation, payment and debt tracking, and dynamic PDF invoices, with role-based access for admin, service, corporate and client users.
Fifth-year engineering student in Artificial Intelligence & Data Science at ESI-SBA, with training in machine learning, deep learning, NLP and large-scale data processing, and hands-on work in forecasting, computer vision and graph learning. I care most about the last mile: carrying a model through to a working application. Interested in applied ML, retrieval-augmented generation and data-intensive systems.