Umer Mansoor (@umer111k) header
Umer Mansoor (@umer111k) profile picture

On a mission to transforms healthcare. CTO @ January AI.

San Francisco Bay Area, CA
Joined July 2025
Photo by Umer Mansoor (@umer111k)Photo by Umer Mansoor (@umer111k)Photo by Umer Mansoor (@umer111k)Photo by Umer Mansoor (@umer111k)Photo by Umer Mansoor (@umer111k)Photo by Umer Mansoor (@umer111k)Photo by Umer Mansoor (@umer111k)Photo by Umer Mansoor (@umer111k)Photo by Umer Mansoor (@umer111k)
Couple of learnings from my 2nd iteration of (local) software factory: 1/ Shared project goals/requirements are a must. 2026 AI writes code that’s good, fast and cheap. But if different people have different beliefs or their own understanding of vague goals, good code in the wrong direction makes matters worse. Engineers can use different harnesses, tools, as long as everyone is working from the same set of project goals. 2/ If the AI agent that’s implementing code cannot answer a design question confidently from goals and requirements, make it stop and ask, not invent its own spec or go with plausible guesses.
1
25
Regarding 1/: in the pre-AI era, features took weeks or months and that slowness gave us time to notice when someone was heading the wrong direction. Tech specs that took days to write got everyone on agree on goals/requirements and create shared understanding. Today engineers move fast and work on several features concurrently. This problem existed before AI; what changed is how fast it compounds. By the time you notice the drift usually bcs your gut tells you something is wrong or a customer complains, a lot of new code is built on the wrong assumption and that wrong assumption became was treated as truth by all future agents.
13
Learning: earlier, I’d allow some ‘sloppy’ PRs if they were low risk or a small feature. Wrong, must actively prevent slop from getting to the codebase. Every sloppy PR, code comment, test code, MD file, becomes the context for future agents. Slop compounds, misleading and confusing future agents leading to poor outcomes, bad design choices, and more slop in the future. A good harness can help a lot in identifying and preventing.
1
25
Umer Mansoor retweeted
We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks: Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better: Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better: Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better: Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work! In summary: - As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding. - Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.
1,711
6,771
1,585
56,867
8,238,788
More is expected from managers now (IC work, research, strategic planning, customer calls, etc) but the primary goal remains unchanged: raise their org’s output. What’s easier now with AI is to process mountains of evidence (real org data) to find bottlenecks, capacity constraints, opportunities to coach, delegate, improve processes and ways to increase leverage.
18
The idea to involve everyone in decision making sounds like it’s creating transparency and building consensus but it’s not. What it often does is that it exports debates and confusion to everyone and hurts more esp. when speed and clarity matters. Also some people get highly invested in their ideas and can’t ‘disagree and commit’. At best, the result is compromise, delivered late. Transparency is not the same as consensus. Leaders should definitely listen, consider different perspectives and then decide with enthusiasm. Explain. It should not feel democratic. Own it.
19
I’m blocking 30-60 minutes each week to write code by hand like we did in the Stone Age. No AI, just the code, IDE with AI features disabled and… Google search.
22
Umer Mansoor retweeted
Alibaba open-sourced the code reviewer it says has served tens of thousands of developers and found millions of defects internally. Open Code Review does not leave the whole review loop to an LLM: deterministic code handles file coverage, bundling, rule matching and comment positioning, while the agent handles reasoning and repository context. On its benchmark of 200 PRs across 50 open-source repos, Alibaba reports higher precision and F1 than Claude Code with the same model, using roughly 1/9 the tokens, with lower recall as the trade-off. GitHub Repo: github.com/alibaba/open-code…
50
183
20
1,540
130,208
🐐 🫡
The actual GOAT of our IT industry who co-created the C programming lang and helped develop the Unix OS. Happy birthday to a true legend whose work shaped modern computing.
1
53
Yuka, but for blood sugar 🩸📈 Scan a grocery barcode → see its predicted glucose spike → get a lower-spike swap from the same shelf. Glucose predictions come from a model trained on thousands of people with CGM, not an LLM's guess. Interested? Comment and will send you the app + 1000 free scans! (not medical advice)
1
1
94
Umer Mansoor retweeted
Sorry, but asking Jacob about AI extinction risk is like asking your AC guy about climate change. Not saying it's necessarily uninteresting or wrong per se but let’s keep things in perspective and hear from the full range of expertise across the ecosystem!
346
663
141
7,029
868,508
Here's a trick for finding bugs and DX gaps in your REST APIs: have an AI one shot a consumer app using the APIs. Do not tell the AI that the goal is to find bugs or it'll start optimizing for that. Just ask it to build the app. When done, ask for every issue it hit and explicitly ask for all the workarounds that it built bcs it has silently patched things behind the scenes, added hack code just to get the job done. That's your bug list. I did this last night with GPT-6 Astra on a few one shot demo apps and it found some hidden issues that tripped it up. I got the list, switched to Opus 5, asked it to verify every single issue independently and then file ticket with details. Result: fewer bugs, better DX.
1
1
44
Astra's one shot Cal AI clone. The UI is 🔥 Total build time: 32 minutes ~$40 (est). App's source code here: github.com/umermansoor/ai-ca…
Claude Opus 5 one shot a Cal AI clone and it COOKED. Snap a meal -> calories, macros and a blood sugar forecast before you eat I literally just asked it to build a clone using January AI APIs. I expected to find several issues, not a working app. Build time: 67 minutes React Native source: github.com/umermansoor/ai-ca…
4
2
3
25,843
Astra's initial version also worked but I did do 2 passes to make some small adjustments
5
meal scan side by side
7
onboarding side by side
10