A glass sprint board of translucent cards being measured by a sweeping cyan light beam

AI Task Estimation

Software & SaaS
Contact Us
Software & SaaS

AI Task Estimation

Manual task estimation was consuming project managers' and team leads' time and producing inconsistent planning data. RemSoft.Dev researched, prototyped and deployed an AI task-estimation and sprint-planning system for a company with more than 100 developers. The model was initialised on a company-specific dataset - real historical employee data - and continuously refined after each completed task; a fine-tuned local model (Phi-3 → Phi-4, 8B parameters, ONNX) decomposes a natural-language project description into tasks and estimates timelines, effort and cost within minutes. After six months in production, estimates reached nearly 90% agreement with the company's strongest developers. Sprint planning dropped to a few hours for project managers, team leads were released from routine estimation work, and their measured performance rose by more than 10%.

AIEstimationSprint planningLocal LLM

Key Challenges

Client: Software company, 100+ developers

Model: Fine-tuned Phi-3 → Phi-4 (8B), ONNX, local

Time in production: 6 months of continuous learning

Integration: Azure DevOps task flow

The project faced several key challenges:

  • Estimates people trust: The bar was the company's best engineers - an estimate that senior developers reject is worthless in planning.
  • Learning from every task: The system had to keep improving from completed work, not freeze at deployment.
  • Private and cheap to run: A fine-tuned 8B local model kept company data in-house and inference costs negligible.

Results

  • ~90% agreement with senior estimates
  • Sprint planning in hours, not days
  • Team-lead performance +10%
  • Local fine-tuned model - private data

~90%

agreement with the strongest developers after six months.

Hours

for sprint planning instead of days.

+10%

measured team-lead performance once freed from estimation.

Related case studies

Want to Build Something Similar?

Tell us about your project - we will propose an approach, a timeline and a first estimate.

Contact Us
An unhandled error has occurred. Reload 🗙