Computer Science @ BMSCE · Open-source builder · AI infrastructure & systems
I build software that remembers, reasons, and holds up outside a demo.
My work sits at the intersection of AI agents, retrieval, developer tooling, and dependable backend systems. I am especially interested in software that preserves context over time, exposes why it made a decision, and gives people a clear way to inspect or correct it.
Right now, that means building local-first memory infrastructure for AI agents, measuring retrieval systems instead of hand-waving about them, and contributing fixes to production open-source projects.
Flagship project — Waggle
Persistent, graph-backed memory for AI agents across sessions and tools.
Waggle is a local-first external memory layer for MCP-compatible agents. It stores decisions, reasons, contradictions, and changing project state as a knowledge graph so a new session can continue from durable context instead of starting from zero.
- graph and hybrid retrieval with temporal and supersession-aware memory
- SQLite by default, optional Neo4j, and local embeddings with no API key required
- MCP integrations for Codex, Claude, Cursor, Gemini CLI, Antigravity, and more
- Graph Studio for inspecting provenance, retrieval, and memory lineage
- portable
.abhigraphs, cross-platform releases, a VS Code extension, and a Codex plugin - benchmarked on LongMemEval with reproducible evaluation artifacts
pipx install waggle-mcp
waggle-mcp setup --yes
waggle-mcp doctorRepository · PyPI · Live Workspace · Releases
| Project | What it does | Built with |
|---|---|---|
| Waggle SemCache | Local-first semantic cache for LLM, RAG, and agent workloads with safe reuse gates, bounded SQLite storage, and inspectable match decisions. | Python, SQLite, MiniLM, OpenAI/LangChain/LiteLLM integrations |
| Waggle Recover | Auditable revenue-recovery agent that separates retrieved context from current authority and keeps deterministic policy in control of every action. | Python, FastAPI, React, LangGraph, Razorpay Test Mode, n8n |
| BorderBridge | Humanitarian identity and case-management platform with explainable confidence scoring, evidence graphs, and self-service refugee workflows. | React, TypeScript, FastAPI, Supabase, custom graph visualization |
| IncidentResponseEnv | Deterministic OpenEnv benchmark for SRE agents investigating and remediating cascading production incidents under a bounded action budget. | Python, FastAPI, Pydantic, Docker, Hugging Face Spaces |
| OCR Q/A Segmentation | End-to-end handwritten exam digitization with OCR, CRF sequence labeling, multi-page stitching, and structured Q/A extraction. | Python, TrOCR, CRF, OpenCV, Gradio |
I enjoy the part of open source where the bug is real, the codebase is unfamiliar, and the fix has to survive production review.
| Upstream | Merged contribution |
|---|---|
kubernetes/kubernetes |
Prevented a CEL composition data race by deep-copying mutable MapType state |
kubernetes/minikube |
Fixed Podman volume mounts on macOS with host.containers.internal support |
kubernetes/minikube |
Corrected nested-VM detection on macOS and removed unnecessary timeout inflation |
containernetworking/cni |
Repaired a broken specification link in the CNI documentation |
- Founded and maintain Waggle, now a community project with 60+ contributors, 900+ commits, PyPI releases, and selection for GirlScript Summer of Code and Social Summer of Code.
- Landed four merged upstream contributions across Kubernetes, Minikube, and the Container Network Interface project.
- Won 1st place among 50 teams at the BMSCE Mega Hackathon with BorderBridge, built in one day.
- Presented Waggle at the Magicball Agentic Summit 2026 to founders and engineers working on AI infrastructure.
- Serve as an AI mentor for the IEEE SIG at BMSCE, teaching LangChain, RAG, agent workflows, and LangGraph.
Languages Python · Go · Java · TypeScript · JavaScript · C
AI systems MCP · RAG · knowledge graphs · embeddings · LangGraph · OCR · CRF
Backend FastAPI · Node.js · REST APIs · auth · SQLite · PostgreSQL · MongoDB
Frontend React · Vite · Tailwind CSS · data visualization
Infrastructure Docker · Kubernetes · Minikube · GitHub Actions · Linux
Engineering testing · benchmarks · retrieval evaluation · release automation
- memory systems that preserve authority, provenance, and change over time
- evaluation methods for long-context retrieval and agent behavior
- distributed systems and Kubernetes internals
- local-first AI products with transparent human control
If you are working on agent memory, AI infrastructure, developer experience, or systems-oriented open source, I would love to compare notes.