Timothee Issenmann

Time to re-introduce myself 👋 I'm Tim, I'm soon graduating from my CS master's and looking for opportunities in SF! I: - founded a photography platform with over 10,000 users, 45,000+ posts and 100s of paying members - built 20+ of the most crucial internal tools at EF across fund, operations, demo day, investor relations, etc - am currently doing research with Codespeak on optimising AI browsing agents for e2e testing if you’re a founder building an ambitious AI product and need someone who can move between customers, product and code, I’d love to talk. DMs open! timissenmann.com/

1 month in sf

a bunch of tiny ui improvements to my web app! (video made in a few prompts with temple.so)

got a demo of what Ben is building and can’t stop thinking about it since

won 2nd place at the forward deployed engineer hackathon at stanford! gdpr right-to-be-forgotten requests are a pain. every request have to be processed manually and coordinated across multiple teams. so this weekend, my team and I built an agentic workflow that crawls through a company's internal systems (think salesforce, gmail, slack, notion...) and redacts/deletes relevant records, keeping a human in the loop for confirmation. thanks to the judges and my amazing team!

I may have accidentally created the most anti-slop corner of the internet. A few years ago, I launched a photography platform called Newgrain, where users are only allowed to post film photos. Instagram was turning sloppy as it launched Reels and turned into TikTok-style feeds. Now with AI, the slop is spreading even more rapidly to every platform. On Newgrain, it’s the opposite. Every day, hundreds of beautiful photos are posted by film shooters who are intentional about their craft. Over 10k users and 50k posts later, it’s become a space to share your best work with no distractions - no ads, no videos, no algorithm. If you love beautiful photos, or if you’re just tired of doomscrolling X, get your dose of inspiration on newgrain.app. I promise it’s worth it.

AI can now build apps fast. But checking whether they actually work as intended is still costly. For my Master’s thesis at Imperial College London, I spent the summer developing Semtag: a semantic representation layer for autonomous testing agents. Semtag embeds hints directly into front-end code, describing each interface element’s role, identity and state, then extracts them into compact snapshots for the agent. I then built a custom agent harness and benchmarked Semtag across 840 runs, three AI models (Claude Haiku 4.5, Sonnet 5, and Opus 5) and 14 web apps.  The result: semantic snapshots were 69% smaller on average than accessibility snapshots, reducing API token cost by up to 16% – without reducing bug-detection performance. A huge thank you to my supervisor, Dr Robert Chatley, and to @abreslav and @codespeak_dev for their close collaboration and support throughout the project! Read the full thesis here: timissenmann.com/thesis.pdf

Are humans still better than AI at improving AI? In most cases no. I joined @apturaai and @OpenAIDevs’ hackathon in @UseCorgi café this weekend in London: The goal: beat a self-improving AI loop on 4 benchmarks in under 12 hours. Each team was allowed to submit one agent skill per benchmark to optimise the performance of GLM-5.3 on them. In 3 of the benchmarks, the recursive self-improving AI beat all human teams. My biggest takeaway: overcomplicated skills perform worse than simple ones. Beyond basic instructions on logging results and approaching the task, making the skill too specific and restrictive made the agent WORSE at solving tasks. Anyone building a harness today should remember: frontier models perform best when left free to decide how to approach a problem.

Pairing this with WebMCP will be insane!

was thinking the same last week! so I built ChatJipity: a clone of the original ChatGPT, powered by GPT-3.5 chatjipiti.timissenmann.com/ enjoy 🫶

When will Anthropic release Haiku 5? It's been almost a year since Haiku 4.5 was launched. GPT 5.6 Luna smashes it now and is much cheaper.

While everyone is waiting on Astra… anyone else miss the original ChatGPT? I do. So I rebuilt it. Introducing ChatJipity: ChatGPT from the good old days, powered by GPT-3.5. enjoy 🫶 chatjipity.timissenmann.com

So many insane tech acquisitions in the last 3 months: - Cursor → SpaceX ($60B) - OpenRouter → Stripe (~$8B) - Hugging Face → Nvidia ($13B) - Airtable → Bending Spoons ($1.3B EV) - GoPro → Starman Optical ($285M) what explains this? why did they all happen in such a short period?

free idea for your portfolio: a dark/light mode toggle for your mockups. Keep the device frames in the same place, just change the screenshots, and your users will feel like they're actually changing the apps' theme!