Tirth Kanani (@tirth_8205) header
Tirth Kanani (@tirth_8205) profile picture

AI/ML engineer | Building smarter systems, debugging chaos, and sharing the journey | Code, ideas, and lessons from the frontlines of machine learning

London, England
Joined August 2019
Photo by Tirth Kanani (@tirth_8205)Photo by Tirth Kanani (@tirth_8205)Photo by Tirth Kanani (@tirth_8205)Photo by Tirth Kanani (@tirth_8205)
Looking forward to it!
accelerating the rate of scientific discovery is the most important thing we can do for the future of humanity
1
34
The beauty 🤌🏻
This Post is from a suspended account. Learn more
1
53
Yesterday I got to demo code-review-graph on stage at Claude Founder House London 🧡 The idea in one line: code-review-graph turns your codebase into a live map, so your AI coding agent reads only the code that matters instead of re-reading the whole repo. Ask an agent, "How does FastAPI build a route handler?" Grep hands the model all 4,912 lines of routing.py. The graph knows where that function starts and ends, so it returns 8. Median 8.9x fewer tokens to reach the same answer across 18 questions on 6 real repos. Open source, runs locally, and plugs into Claude Code and most other coding agents over MCP. Special thanks to @lamismukta and the @AnthropicAI @claudeai team for a brilliant day. The "Built with Claude" wall said it all: a room full of people shipping real things. ⭐ github.com/tirth8205/code-re…
8
1
19
277
Tirth Kanani retweeted
We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so. Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training. You can read the full post here: darioamodei.com/post/we-must…
10,538
16,295
11,365
87,525
77,140,392
“Unsolved” just got a shorter shelf life.
Replying to @OpenAI
GPT-6 Astra is state-of-the-art on FrontierMath Tier 4, ARC-AGI 3, and TerminalBench-4.0. GPT‑6 Astra is also a major advance for scientific discovery, with state-of-the-art performance on Terminal-Bench Science 0.1 and HealthBench Pro.
1
67
Tirth Kanani retweeted
the next decade will be a lot about specifying what we want matter to do - and computationally discovering how to make it. alchemy tried to program matter. chemistry taught us the rules. AI may finally give us the compiler.
2
2
1
71
3,197
Tirth Kanani retweeted
Forget FAANG Forget MANGO The whole world is run by ONLYFANS now.
233
1,719
197
16,341
938,623
Tirth Kanani retweeted
This Independence Day 🇮🇳, I asked, "What will the next century of young Indians look like?" $0.01 Drone Deliveries. UAVs that never land. Groceries in under 10 minutes. Meet The 22nd Century Indian. A Documentary on a New India.
412
1,290
463
6,340
1,403,038
Today’s GitHub Trending Top 5 was a nice mix: auth, Java infrastructure, agent tooling, code graphs, and video editing for coding agents. Shoutout to @melizeche @cpovirk @SigGravitas @tirth_8205 @mamagnus00 @gregpr07. #GitHub #OpenSource #AI
1
3
99
Tirth Kanani retweeted
AI code review tools hallucinate context because they don't actually know your repo. code-review-graph by @tirth_8205 builds a local graph of your codebase so your AI assistant gets blast-radius analysis, call graphs, and risk scoring — not guesses. Open source. Full breakdown next post. #DevTools
1
1
2
67
Today’s GitHub Trending list was a good mix instead of another all-agent-tools roundup. I put together a 15s scan featuring @tirth_8205, @KVCache_AI, @yourtowhid, jellium-desktop, and @OpenAtMicrosoft, covering code context, LLM inference, local research, desktop apps, and terminal tooling. #GitHubTrending #OpenSource #DevTools
1
4
156
Tirth Kanani retweeted
Fable 5 is back.
4,234
10,969
5,015
89,923
17,360,336
😂😂
Me using Claude Opus 4.8 to rename a file
2
2
141
code-review-graph v2.3.5 just shipped. github.com/tirth8205/code-re… Headline feature: real-time token savings, visible to humans. When you run `code-review-graph detect-changes --brief` or `update --brief`, the CLI prints a boxed panel showing exactly how much context the graph kept out of your AI agent's window.
1
2
103
Don't trust an estimate? Add --verify. It cross-checks the same calculation against tiktoken's cl100k_base (GPT-4 family) and grows a "Verified (tiktoken)" row in the panel. Calibration across 222 source files: the chars/4 estimate stays within +0.5% of real tokens in aggregate. Per-repo bias is bounded to ±12%, and the ratio holds because both sides are equally biased.
1
1
38
Also in 2.3.5: - Deterministic eval pipeline (pinned SHAs, seeded Leiden, full clones) - Multi-hop retrieval benchmark — 0.909 average score - Richer semantic search (compound NL queries land on the right node) Full notes: github.com/tirth8205/code-re… Reproduce: docs/REPRODUCING.md
2
49
A huge thank you to everyone who’s been using and contributing to it! We’re thrilled to announce that a lot of new features are on the way!
Replying to @tirth_8205
@tirth_8205 just crossed 17k stars on code-review-graph 🔥 It’s a local knowledge graph that lets coding agents read only the files an edit touches. Up to 49x token reduction on real repos. Built solo in London. 👇
2
133
Open call. I need one person to build something with me. Not a co-founder, not staff. Someone who reads the footnotes and builds by feel. Anywhere in the world. The rest is here: tirthkanani.com/one
1
1
84