👋 About Me

I’m Oussama Gabouj, co-founder and CTO of Compresr (Y Combinator W26), where we build context-compression infrastructure for LLMs. I hold an MSc in Data Science from EPFL with a minor in Cyber Security.

My work is about making language models behave under tight context budgets — first as a researcher at EPFL’s dlab, where I was first author at EMNLP 2025 Findings, and now as the person who builds and runs the systems that put that research into production: the API and billing platform, an open-source Go proxy, the GPU serving layer, and the benchmark harness we use to decide whether a change is real.

I also keep an Engineering & Research Journal — write-ups of bugs and measurement failures worth remembering: a GPU that silently ran on CPU, an autoscaling study that crowned a policy which never scaled, a benchmark headline that didn’t survive correct cost accounting.

🚀 Currently

We spun out of EPFL’s dlab and were taken into Y Combinator W26. I’m the infrastructure engineer on a four-person team, which in practice means:

• Context Gateway — an open-source Go proxy (637 ★, Apache-2.0) that sits between a coding agent and its LLM and keeps the context window small. It summarises in a background worker and intercepts the agent’s own compaction request, so compacting returns something that already exists instead of stalling the session. Works with Claude Code, Codex, Cursor, OpenCode and OpenClaw.
• The cloud platform — a multi-tenant API with Postgres row-level security, Redis buffering and Stripe billing, plus an on-premise product for customers who can’t send data out.
• GPU serving — our compression models served under vLLM on EKS, autoscaled by a custom CloudWatch metric.
• The benchmark harness — 12 long-context suites, plus a sandboxed setup for running coding agents under compression.

In production, compression has cut customer inference costs and latency substantially while leaving answer quality intact — measured against parity gates rather than asserted.

📄 Selected Publications

• GRAD: Generative Retrieval-Aligned Demonstration Sampler for Efficient Few-Shot Reasoning — first author (equal contribution), EMNLP 2025 Findings. Instead of retrieving few-shot examples from a database, we train an LLM with GRPO to write them per question.
• Cmprsr: Abstractive Token-Level Question-Agnostic Prompt Compressor — co-author, arXiv:2511.12281, under review at ACL 2026.
• Generative approaches to kinetic parameter inference in metabolic networks — co-author, Nature Communications, 2026.

🎓 Education

MSc in Data Science, Minor in Cyber Security — GPA 5.41/6
Master Thesis: Structured Representations for Fine-Grained Text-to-Image Retrieval in Remote Sensing (Prof. Devis Tuia, EPFL ENAC — hosted by AXA Group Operations)
EPFL, Switzerland | 2023 – 2025

BSc in Microengineering — GPA 5.33/6
EPFL, Switzerland | 2020 – 2023

🏢 Labs & Industry Experience

Research Labs @ EPFL

• DLab: The Data Science Lab focuses on transforming large-scale data into meaningful insights by developing algorithms in natural language processing, machine learning, and computational social science. I spent a year there as a research assistant with Prof. Robert West.
• LCSB: The Laboratory of Computational Systems Biotechnology specializes in reconstructing and analyzing biological networks to understand cellular processes through computational models.
• DISAL: The Distributed Intelligent Systems and Algorithms Laboratory develops methodologies for distributed, intelligent systems, emphasizing cyber-physical systems like multi-robot systems and sensor networks. I joined on a competitive Summer in the Lab fellowship.

Industry

• Compresr Inc.: LLM context-compression infrastructure. Y Combinator W26.
• AXA Group Operations (Switzerland): The IT services division of AXA, focusing on creating innovative technology and data solutions to support AXA’s ambition of being a customer-focused, tech-led company.
• Pixalione (Paris): A digital marketing agency specializing in SEO, Paid Media, and Data Analytics, combining human expertise with proprietary algorithmic tools to optimize web presence.

🏆 Honours & Awards

• Y Combinator W26 — selected for the Winter 2026 batch.
• Venture-backed pre-seed raised for Compresr as technical co-founder.
• Summer in the Lab Scholarship, DISAL, EPFL — a competitive fellowship funding a full-time summer of research (2023).

🛠️ Technical Skills

• Languages: Python, Go, TypeScript, SQL, C/C++, Bash
• ML & LLMs: PyTorch, HuggingFace Transformers, TRL, PEFT/LoRA, vLLM, reinforcement learning (GRPO, PPO, reward design), DPO, quantisation
• LLM systems: LangChain, LangGraph, LlamaIndex, LiteLLM, MCP, agent tooling, RAG and long-context evaluation
• Backend: FastAPI, PostgreSQL (row-level security), Redis, Stripe, REST/SSE streaming APIs, Next.js, React
• Cloud & infra: AWS (EKS, ECS Fargate, SageMaker, ElastiCache, ECR, ALB, CloudWatch), Terraform, Docker, Kubernetes, Karpenter, GitHub Actions
• Practices: pytest / Vitest / Playwright, CI/CD, observability (Sentry, Prometheus), container security & SBOM, SOC 2