Projects
Production systems, research prototypes, and open-source tools
⭐Featured
HyperSentry
An agentic security copilot that uses policy checks + retrieval to generate remediation PRs and accelerate incident analysis.
- False positives under 3%
- Auto-merged 30%+ remediation PRs
SchemaPulse
OpenAI-compatible inference layer for schema-valid structured outputs with streaming tool-call parsing.
- Streaming incremental tool-call parsing
- Adaptive guided decoding for schema validity
EdgeJury
Multi-LLM council with cross-review for truthful QA on serverless edge inference.
- Cross-reviewed small-model ensembles
- Structured critique + synthesis
Ghosts in the Scale
Prompt injection via image resampling & certified defenses for multimodal systems.
- Black-box downscaler fingerprinting
- ScaleJail-mini benchmark
All Projects
TariffWhisperer
End-to-end RAG system over CBP rulings for explainable HTS code recommendations.
- Dense + BM25 hybrid retrieval
- Sub-500ms retrieval latency
Adversarial Phishing Detection
Simulated arms-race where an LLM attacker and classifier defender co-evolve.
- LLM-powered phishing generation
- Continuous adversarial training
Lookahead RAG
Speculative retrieval planning for low-latency multi-hop QA.
- Retrieval dependency graph planning
- Parallel retrieval execution
BudgetBench
Adaptive reasoning-budget control for cost-efficient LLM inference.
- Accuracy–cost Pareto curves
- Adaptive budget escalation
PaperChatbot
Structure-aware + verifiable/abstain RAG for research papers.
- Section-structured retrieval
- Every-claim-must-be-cited mode
AdapterLLRD
Memory-free multilingual continual learning with PEFT + LLRD.
- 50+ hop sequences without replay
- Privacy-compatible (no buffers)
OOD-Aware Fairness
Selective classification for toxicity detection with fairness constraints.
- OOD score inversion diagnosis
- 86% reduction in FPR gap