Skip to content
View 0xA8hinav's full-sized avatar
🦝
Hi
🦝
Hi

Highlights

  • Pro

Block or report 0xA8hinav

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
0xA8hinav/README.md
Typing SVG


About Me

I am a 3rd year B.Tech Computer Science student at KIET Group of Institutions. My technical work focuses on three primary areas:

  • Agentic AI - MCP and WebMCP tooling, LLM orchestration, and memory systems for agents.
  • Backend Systems - APIs, message queues, worker processes, and databases required for agents to execute tasks deterministically.
  • DevOps and Systems - Docker, Kubernetes, CNCF runtimes, and distributed systems reliability (AgentCube, llm-d).

I believe that an AI agent is only as reliable as the infrastructure supporting it.

  • Currently building: An automated dubbing pipeline for Indian languages using Go and message queues.
  • Currently studying: Agent evaluation benchmarks, Kubernetes internal architecture, and real-time streaming audio.
  • Open source contributions: CNCF AgentCube, llm-d, better-auth, Pangolin

Featured Projects

Agentic AI / Edge

TypeScript - Next.js - WebMCP - WebCodecs

A browser-based screen-recording studio where AI agents collaborate directly on the same canvas (live demo). It contains 30 WebMCP tools that allow an AI agent to inspect the canvas and suggest zoom actions, cuts, and subtitles as staged changes for user approval. Accurate synchronization is maintained with dual time-space mapping and snapping logic. Video decoding, rendering, and exporting run fully inside the browser using WebCodecs and OPFS at 4K 60fps without uploading files to external servers. Built for the OpenAI WebMCP Challenge.

Agentic AI / Backend

TypeScript - Electron - Playwright

A web monitoring tool where users select any webpage element and specify a condition, such as "alert me when price < INR 2,000." An LLM parses the user request once into a deterministic XPath selector and a fallback anchor. Following this, a background worker using Express, BullMQ, and Playwright executes scheduled checks continuously without recurring LLM API expenses.

Agentic AI / Backend

Python - Audio ML - LLMs

An end-to-end speech processing pipeline: custom noise filtering -> speaker diarization (pyannote and ECAPA-TDNN) -> automatic speech recognition (ASR) -> emotion recognition using speaker-normalized acoustic and prosodic features implemented from research literature -> automated compliance scoring and CRM summary generation using LLMs.

Agentic AI / Edge

TypeScript - Edge AI - Vector Search

A unified memory system using the embedded Gemini Nano model in Chrome and local vector search to share context across ChatGPT, Claude, and Gemini without sending user data to external cloud services.

Other projects: Fraktal - no-code workflow engine using Redis pub/sub and runtime state injection | INSIGHTS - AI curated news feed with summaries and scheduled email digests | SunoHQ - voice agent platform for Indian regional languages using Sarvam STT/TTS | Lumio - autonomous content moderation agent running on Arweave and AO


Open Source Contributions (CNCF & Major Projects)

CNCF AgentCube (Volcano) - Sandboxed Runtime for AI Agents (6 Merged PRs)
Contributor - Go - May 2026
Resolved concurrency issues, improved garbage collection handling, and enhanced sandbox lifecycle reliability for the agent session runtime.
  • #339 feat(picod): Added gzip middleware for response compression.
  • #321 fix(workloadmanager): Used detached context for store deletion and added missing log output.
  • #338 fix(router,workloadmanager): Populated sandbox Kind and Namespace fields during creation.
  • #335 fix(workloadmanager): Deduplicated garbage collection candidates before deletion to avoid race conditions.
  • #311 fix(sdk-python): Corrected router port number and fixed session race during initialization.
  • #351 fix: Code style and linter cleanup.
CNCF llm-d - Kubernetes-Native LLM Inference (4 Merged PRs)
Contributor - Go - Aug 2026 - Repositories: llm-d, llm-d-async
Contributed to the durability and monitoring features of the asynchronous batch processing pipeline. All modifications were verified against Redis 7 with unit, integration, and race detection tests (go test -race).
  • #412 feat(redis) (+1,744 lines): Implemented reliable dequeue functionality using lease claims for the sorted-set transport. Replaced destructive ZPOPMIN operations with at-least-once delivery semantics using atomic Lua scripts, token fencing against stale owners, periodic heartbeats, background reclaim routines, and failure recovery. Enqueued requests survive process termination, out-of-memory errors, and node failures. Fixes #404.
  • #400 feat(metrics) (+847 lines): Implemented deadline proximity histogram (llm_d_async_async_deadline_proximity_millis) to alert when queued batches approach expiration. Used a single Redis TxPipelined call for accurate bucket updates with minimal polling overhead. Fixes #345.
  • #401 feat(metrics): Added token throughput counter (llm_d_async_async_tokens_total with input and output labels) by parsing OpenAI usage values in 2xx responses to assist with inference capacity planning. Fixes #344.
  • #2306 docs: Added Prometheus metric references for the router across official documentation guides. Fixes #1777.
better-auth - TypeScript Auth Framework (1 Merged PR)
Contributor - TypeScript - Feb 2026
PR #8177: fix(client): broadcast session updates to other tabs on sign-out and user update
  • Integrated broadcastSessionUpdate into atomListeners for /sign-out, /update-user, and /update-session.
  • Added an optional callback argument to ClientAtomListener and added 2 unit tests to verify cross-tab broadcast behavior.
Pangolin - Self-Hosted Tunneled Reverse Proxy (2 Merged PRs)
Contributor - TypeScript - Feb 2026
PR #2535: fix: correct session DELETE tautology and HTTP cookie domain interpolation
  • Fixed an invalid DELETE query condition that cleared the entire resourceSessions table when a single session expired.
  • Corrected cookie string interpolation that caused web browsers to reject authentication cookies.

PR #2537: fix: sync resource toggle states with context on initial load
  • Fixed an issue where rule toggles, SSO, and email whitelist settings defaulted to disabled state upon initial page render.

Achievements

  • Winner - Surreal World Asset Buildathon (June 2025) - Decentralized digital rights infrastructure
  • Finalist - Agents of the Permaweb Hackathon (August 2025) - Autonomous moderation agent on AO/Arweave
  • Finalist - World Computer Hacker League by ICP (September 2025) - No-code Web3 and AI workflow automation
  • Finalist - PSB FinShield Hackathon, IIT Hyderabad (September 2025) - Behavioral credit risk prediction using GNN (95.28% R^2)

github-snake

"I build AI agents and the underlying infrastructure that runs them."

Pinned Loading

  1. aether aether Public

    TypeScript

  2. SVAR SVAR Public

    SVAR (Sentiment and Voice Analytics for Real-time QA) - An end-to-end Hindi call analytics pipeline combining acoustic denoising, speaker diarization, and MuRIL-based sentiment analysis to auto-sco…

    Python

  3. fraktal-app/fraktal fraktal-app/fraktal Public

    No-code automation on-chain. Build powerful workflows on the Internet Computer.

    TypeScript 2 2

  4. llm-d/llm-d llm-d/llm-d Public

    Achieve state of the art inference performance with modern accelerators on Kubernetes

    Shell 4.5k 753