Agentic loop planner-executor for reliable multi-file edits BC-964

Genre de projet: Innovation
Discipline(s) souhaitée(s): Génie - informatique / électrique, Génie, Informatique, Sciences mathématiques, Mathématiques
Entreprise: Farpoint Technologies Inc.
Durée du projet: 6 mois à 1 an
Date souhaitée de début: Dès que possible
Langue exigée: Anglais
Emplacement(s): Vancouver, BC, Canada
Nombre de postes: 2
Niveau de scolarité désiré: MaîtriseDoctorat
Ouvert aux candidatures de personnes inscrites à un établissement à l’extérieur du Canada: No

Au sujet de l’entreprise: 

Farpoint Technologies is a leading AI digital transformation consulting company that empowers top-tier organizations to build the AI-assisted workforce of the future. We specialize in providing consulting services to large public, private, and government entities, helping them create AI-accelerated workflows and innovative solutions. Our expertise includes leveraging LLMs, diffusion models, multimodal models, and executing special projects.

Veuillez décrire le projet.: 

This project develops an agentic loop that delivers non-deterministic but reliable, multi-file code edits through structured collaboration between a large model and one or more smaller models. Unlike single-model approaches, this system clearly delineates planning and execution phases via detailed JSON-based task specifications, ensuring improved consistency, clarity, and reliability in automated software development workflows. The innovation significantly reduces ambiguities, enabling precise and parallelized multi-file transformations while ensuring accurate, reproducible code modifications.

Main tasks for the candidate:
• Define comprehensive JSON schemas for specifying and communicating multi-file code transformations.
• Implement an asynchronous dispatcher distributing file-specific instructions to multiple parallel executor models.
• Develop a structured feedback mechanism allowing executors to request clarifications from the planner.
• Create a conflict-free merge module ensuring deterministic integration of concurrent edits.
• Benchmark the system's reproducibility and reliability on established code-generation tasks.
• Stretch goal: New record on SWE-Verified

Methodology / techniques:
• JSON schema definition and evaluation.
• Asynchronous task orchestration and parallel processing.
• Structured executor-planner feedback communication protocols.
• Conflict resolution leveraging CRDT-inspired merge strategies.
• Reproducibility benchmarking and variance analysis using SWE-bench.

Expertise ou compétences exigées: 

Required expertise/skills:
• Expertise in defining and managing structured JSON schemas.
• Familiarity with large language model API integrations and streaming completions.
• Understanding of Git diff management, parsing, and patch application.
• Basic knowledge of CRDT-based conflict-resolution techniques.
• Ability to script reproducible benchmarks (HumanEval, MBPP, or equivalent evaluation suites).

Assets (optional):
• Experience in concurrent task orchestration systems
• Familiarity with commit-signing and auditing best practices.