Bobbili Somanadh
Vasudevara Sasi Sundar

Problem: Contract review quality varied across reviewers and document types.

Approach: LLM pipeline for field extraction and pricing anomaly checks.

Experimentation: Compared prompt styles and validation rules on representative procurement samples.

Results: Produced more consistent extraction outputs and reduced manual correction in repeated runs.

Insights: Post-processing constraints improved reliability more than prompt verbosity.

Problem: Multi-step agent behavior is hard to evaluate without a controlled setup.

Approach: API-integrated agent chaining tools and reasoning steps via FastAPI.

Experimentation: Tested mixed task sets across structured and ambiguous categories.

Results: Better structured-task performance, with clear weaknesses on ambiguous tasks.

Insights: Tool selection policy is a major failure point.

Problem: Resume scoring changed across reviewers and sessions.

Approach: TF-IDF + ML ranking pipeline exposed through Flask API.

Experimentation: Evaluated multiple model variants on labeled resume sets.

Results: Improved ranking consistency versus manual-only review in internal evaluation.

Insights: Feature calibration matters more than model family choice.

Problem: Manual user segmentation produced unstable clusters.

Approach: Feature-engineering and clustering pipeline with stability scoring.

Experimentation: Benchmarked cluster quality across multiple feature sets on a medium-sized user dataset.

Results: Produced more stable clusters and fewer low-confidence assignments than the initial baseline.

Insights: Domain feature design outperformed algorithm switching.