Hacker
Abhinav Ayyagari
@abhinav_ayyagari
Elsewhere
Stack
- TypeScript
- JavaScript
- Python
- React
- React Native
- scikit-learn
- pandas
- NumPy
- TensorFlow
Projects
KAL
No Hands: The Wispr ChallengeOct 2026
Students juggle deadlines, reminders and rough days across five different apps, and none of them listen. Kal is a voice-first life companion for students. You say one messy sentence, and Kal sorts it: deadlines and alarms go into your plan, feelings go into your journal with a gentle reply, and the people and decisions in your life go into a memory you can ask questions of later. When a day is hard, it slows down instead of handing you a to-do list. Kal has a single text box, and you dictate into it with Wispr Flow, hands-free, in your own words. An AI agent (Nemotron 3 Super) reads the meaning and the tone of what you said, then turns it into real alarms, journal entries and memory, all stored on Supabase.
- Top 50
A shared FalkorDB graph that lets AI coding agents work in parallel without colliding.
Graph HacksOct 2026
When several AI agents work on one codebase, they collide: they edit the same function, act on stale assumptions, and produce work that won’t merge. Nodes fixes that with a shared graph. Nodes indexes a repo (14 languages plus git history) into FalkorDB and connects to Claude Code through MCP tools and hooks. Before every Edit, Write or file-writing Bash command, the agent’s action is checked against the graph: Ownership: an atomic lease, so two agents never claim the same function. Waiting and hand-off: an agent that needs busy code waits, and gets the code handed over when the owner finishes. Deadlocks: a wait-for cycle is caught and sent to a human. Facts: a claim is only trusted if it matches its source. Impact: agents see what their change reaches, through call and import paths, before they edit. A live dashboard shows each agent as a cute “node” mascot walking to the code it touches, with every decision, hand-off and Cypher query on screen. The results are pre-registered and held-out, with the same model in every condition. On HiddenBench (40 tasks, Claude Haiku 5.5), groups decide correctly 38% of the time with Nodes, against 12–15% for chat, placebo and an unchecked board (p = 0.006). On CooperBench (45 pairs, Claude Sonnet 5.5), two agents’ work merges cleanly 35/45 times, against 22/45 for the same procedure over chat (p = 0.002). The product replays all 4,916 recorded research tool calls identically, with 0 double-claims at 100 concurrent agents.
Git Outta My Way@mahathi_aishwarya,@padmini21