Vipin Research Project Map
Question
What are Vipin’s current main research projects, and how should the wiki remember their roles without modifying the project repositories?
Summary
Vipin’s current research portfolio has three main technical arcs:
- Decision-relevant uncertainty in recommendation: uncertainty, truce-rec, and tgl-rec.
- Closed-loop uncertainty-aware optimization: analog-agent and protein-optimization-feedback-shift.
- Agent task completion and specification grounding: donebench.
The most important wiki principle is to preserve each project’s claim boundary. Many repos explicitly warn against turning diagnostics, smoke runs, or partial official-baseline wrappers into paper-facing conclusions.
Project Roles
| Project | Role | Current wiki interpretation |
|---|---|---|
| analog-agent | AI4EDA analog design agent | Layered agent architecture for SPICE-grounded analog circuit optimization. |
| uncertainty | Main LLM4Rec uncertainty project | C-CRP and task-grounded calibrated uncertainty under same-candidate evaluation. |
| truce-rec | Uncertainty-aware generative recommendation system | CURE/TRUCE route with strict evidence labels and official-native baseline gate. |
| tgl-rec | Temporal recommendation evidence project | Tests temporal next-need transitions vs semantic similarity/popularity. |
| donebench | Agent benchmark | Specification grounding benchmark with deterministic DoneSpec grading. |
| uncertaintyprotein-ai4s / protein-optimization-feedback-shift | AI4S protein uncertainty project | Feedback-shift study where calibration and decision quality can diverge. |
Cross-Project Pattern
- EXTRACTED: The recommendation projects share the rule that a method is not paper evidence until the same-candidate protocol, score coverage, provenance, and paired/statistical gates pass.
- EXTRACTED: donebench uses deterministic executable grading rather than LLM-as-judge, which fits the same evidence-discipline pattern.
- EXTRACTED: analog-agent and protein-optimization-feedback-shift both treat closed-loop decision quality as more important than passive prediction quality.
- INFERRED: The unifying research taste is controlled decision systems under uncertainty: when to trust a model, when to defer to truth/evidence, and how to make claims only after protocol gates pass.
Immediate Use In The Wiki
- For quick questions about the research portfolio, start from research-projects and this map.
- For project-specific questions, open the relevant entity page first, then use the source paths recorded there.
- For paper-claim questions, prefer the external project’s own canonical docs before relying on old chat memories or generated output folders.
- For research-idea questions, apply research-ideation-policy: do not merely stitch existing project parts together; be willing to propose radical reframing when it creates a stronger claim.
Counterpoints and Gaps
- This pass did not ingest all project PDFs, paper drafts, result tables, or code-level algorithms.
AI for Scienceas a folder did not show visible content in the inspected top-level pass; the active AI4S project appears to be protein-optimization-feedback-shift.Analog-RAPID,analog-uaas, andAnalog-AEDAshould receive deeper lineage notes later if they still influence the current analog-agent paper.Uncertainty-LLM4Recneeds an archival source note if older Week7/Week8 evidence is reused.