Work

• case-computer: live. Case Computer. An AI agent with its own computer: tell it a task, it opens the apps, does the work, and sends you the result. case.computer
• case-sdk: ~99%. Task success driving macOS apps like DaVinci Resolve, against ~70% zero-shot. Agents run typed procedures on the accessibility tree instead of clicking pixels. GitHub
• attention.triton: 50×. Faster than PyTorch SDPA at 64K tokens. Linear attention, rebuilt from scratch in Triton to learn GPU programming. GitHubWrite-up
• agent-debugger: rewind. Chrome DevTools for AI agents. Step back through a real run, fork it at any step, patch the state, replay. GitHub
• cortex: 1 graph. Every ChatGPT, Claude and Perplexity chat in one knowledge graph, each fact linked to its source. Agents read it over MCP. askcortex.app
• after-effects-mcp: 70+. Tools that let an AI assistant build compositions, layers and keyframes in Adobe After Effects. GitHub
• vlm-140m: −7%. Perplexity against a cross-attention baseline. A 140M-parameter vision-language model, built from scratch, that routes tokens by modality. GitHub
• anchor: proof. “Am I eligible for that leave policy?” Anchor turns written policy into formal logic and checks it. Built at QpiAI, open source. GitHub
• proced-mem: 30–42%. A paper I wrote at QpiAI: a benchmark showing agents lose 30–42% when the task is new. MemAgents Workshop, ICLR 2026. PaperCode
• 50x-attention.md: O(n). 50x Faster Attention: what I learned rebuilding linear attention in Triton. Why O(n) finally beats O(n²) at 64K tokens, and where my kernel breaks. Read on Medium
• all-posts.md: more. Everything else I've written, mostly honest notes on experiments, including the ones that failed. Medium

Email ishantkohar@gmail.com. Also on X, GitHub, LinkedIn and Google Scholar.