Skip to content
pushkal.

Projects

Things I've built

Research systems, open-source tooling, and products. The featured set is where most of my time goes; expand to see everything.

VALOR-RAG

Patent PendingHICSS Submission

Trust-scored retrieval defense for RAG pipelines

A security layer for retrieval-augmented generation that assigns dynamic trust scores to retrieved documents, detects corpus poisoning, and adapts its thresholds as attack patterns shift. Developed in collaboration with Washington State University.

  • Dynamic trust scoring over retrieved documents at query time
  • Poison detection that flags adversarial insertions before generation
  • Adaptive thresholds tuned against evolving attack distributions

Architecture

  1. Retrieval

    Vector store query with candidate document set

  2. Trust Scoring

    Per-document trust signal from provenance, consistency, and semantic drift

  3. Poison Detection

    Adversarial pattern classifier over candidate contexts

  4. Adaptive Gate

    Threshold policy that admits or quarantines documents

  5. Generation

    LLM generates only from admitted, trust-weighted context

Python · PyTorch · Transformers · Vector Databases

Meridian

Founder

Operational infrastructure for tutoring businesses

A multi-tenant CRM platform built for session-based education businesses. Generic CRMs model deals; tutoring businesses run on recurring sessions, parent communication, and per-student billing — Meridian models that directly.

  • Authentication and role-based access control across tenants
  • Workflow automation for scheduling, reminders, and follow-ups
  • Notifications and analytics tuned to session-based revenue
  • Multi-tenant architecture with per-org data isolation

Next.js · TypeScript · FastAPI · PostgreSQL

RAGuard

Open-source RAG security toolkit

Prompt-injection and retrieval-attack detection for RAG applications. Ships as a drop-in middleware layer: score incoming queries and retrieved contexts, block or flag adversarial patterns before they reach the model.

  • Prompt injection detection across common attack categories
  • Retrieval-context screening before generation
  • Drop-in FastAPI middleware integration

Python · Transformers · FastAPI

Context Bridge

Persistent memory for Claude Code

A checkpointing and context-restoration system for AI coding agents. Records goals, decisions, blockers, and attempt history across sessions so an agent resumes exactly where it left off instead of rediscovering state.

  • Session checkpointing with goals, blockers, and decisions
  • Cross-session context restoration on startup
  • Stagnation detection and attempt-replay to prevent repeated failures

Python · FastAPI · SQLite

LitMind

AI-native research assistant

A Chrome extension for researchers: semantic annotation over papers, related-work discovery, and a personal knowledge layer that compounds as you read.

  • Semantic annotation directly over paper text
  • Research discovery from what you have already read

TypeScript · React · Embeddings

Margined

LLM unit economics platform

Cost, margin, and profitability tracking for LLM-powered products. Connects inference expenses to features and customers so teams can see which usage is profitable and which is subsidized.

  • Per-feature and per-customer inference cost attribution
  • Margin and profitability dashboards

TypeScript · Next.js · PostgreSQL