Hackathon Portal
AI Tinkerers - Paris
Sign in

Hackathon: Benchmarking Small Language Models in the Real World

Sign in
Engineering Lead at Stealth AI Startup - Stealth AI Startup
Own AI-native architecture and backend pipelines remotely with no equity gimmicks.
All Jobs
Final round winners have been announced. View Results
Team

Polaris

Project Concept

No description has been added yet.

Entry

Status: Submitted

Last saved: April 18 at 5:32 PM CEST

View Entry

Team Roster

Message board not available for this team yet.

Hippolyte Dupont Team Lead RSVP Approved

Ingénier Logiciel & IA at OMNES Education
Prompt engineering and cascade orchestration Designed the system prompt with a modern Polars cheat sheet and 4 targeted few-shot examples, the enriched schema formatting (types + n_rows + a passive reminder of valid column names at the end of the user turn to counter the model's recency bias). Implemented the full cascade (L1→L4) with the static validator, mock execution on empty DataFrames to catch runtime errors without access to real data, and the structured feedback that reformulates Polars errors (ColumnNotFoundError, DuplicateError…) into clear instructions for the retry. Refactored the code into decoupled modules (prompt / model / cascade) and wrote the 85 regression tests.
I am an AI Engineer graduated from CY Tech and IIT in Chicago
Cinema, sports
Improving students resumes with AI

Ghaith ABDESSALEM RSVP Approved

Consultant at ALTEN
Model selection and dataset construction Benchmarked several SLMs (LFM2-8B-A1B, Gemma 4 E2B, Mistral-7B) to guide the final choice. Built the entire dataset pipeline: TPC-H table generation (dataset/gen_tpch.py), construction of validated seeds (data/seeds.jsonl: 25 questions with question / tables / reference_code / expected_output_hash / tags / difficulty), sandboxed executor (dataset/executor.py), canonical hash tolerant to float rounding and column ordering (dataset/compare.py, hashing.py). This evaluation harness locally reproduces the official metric, which let us iterate in seconds rather than in submissions.
RPA Consultant and Python developer based in Paris, with experience building automation solutions for enterprise systems (SAP, web apps, databases). I specialize in designing scalable, production-grade automations using Python & UiPath. Recently, I’ve been focusing on integrating AI into automation by using LLMs and exploring agentic systems capable of reasoning and executing tasks. I’m particularly interested in bridging the gap between traditional automation and adaptive AI systems. Background in engineering and information systems, with a strong focus on practical, real-world impact.
Agentic AI systems, autonomous workflows, and LLM-powered tools. AI for operations and business process automation. Multi-agent systems and tool-using agents. Applied AI with real-world impact (not just demos). Also interested in meeting builders working on GenAI products, early-stage startups, and research-to-product transitions.
I’m currently building AI-augmented automation systems that combine RPA (Python, UiPath) with LLM-based workflows. I’m exploring agentic architectures to automate complex business processes like document understanding, decision-making, and multi-step task execution. I’m also working on structuring scalable automation frameworks (queues, logging, orchestration) with AI capabilities in order to detect technical bugs and automatically resolve them.

Jacques Dumora RSVP Approved

AI Engineer at Tw3 partners
Grammar-constrained generation Built a Lark grammar for modern Polars (dataset/polars_grammar.py) and its integration with Outlines CFG. The grammar is specialized per request: the table and column names from the received schema are injected as terminals, which forces the decoder to produce only valid references by construction. Physically eliminates hallucinated APIs (with_column, pl.desc…) and nonexistent column references when active. Also set up an SFT fine-tuning pipeline (Spider loader + SQL oracle + build_sft) and the CUDA OOM fixes needed to fit memory on the RTX 5090.
I’m Jacques Dumora! Freshly graduated in computer science with a specialization in artificial intelligence, I moved to Paris in September for my first job in this field. I love creating, it’s this passion that drove me to work in a universe where only imagination sets the limits!
GenAI/AI engineering, MCP , Agent development , Software engineers
Grape is a tool that intelligently centralizes a wide range of documentation resources to guide both AI agents and developers in building their applications: database schemas, Swagger documentation, URL structures, and more. Everything is designed to streamline the design process and accelerate the journey from idea to product. In addition, an AI agent assists the developer in making key architectural decisions.