Technical creator. Involved in everything AI, data, and tech
Joined March 2023
- Tweets1.4K
- Following202
- Followers79
- Likes14.2K
Pinned Tweet
The edge computing market hit $168B in 2025. IoT devices are headed to 39 billion by 2030.
And yet: most vector database are mot built for air-gapped or latency-critical environments.
Healthcare. Manufacturing. Military. The cloud can't serve them, and nobody's building for them.
Your favorite tech creator retweeted
Building an AI agent with memory?
Know these 4 types:
1. Working: current task
2. Episodic: past interactions
3. Semantic: learned facts
4. Procedural: workflows
Actian breaks down how they fit into production agent architectures:
dev.to/actiandev/what-you-ne…
Your favorite tech creator retweeted
Your AI agent has a web search tool but still answers current questions from memory?
Check the tool description.
"Search the web" isn't enough. Tell it WHEN to search: new releases, prices, versions, news, etc.
Octen explains this with a working LangChain example:
octen.ai/guides/how-to-add-l…
Your favorite tech creator retweeted
Before connecting a scraping API to your AI agent, check:
- Anti-bot reliability
- JS rendering
- JSON/Markdown output
- Request latency
- Cost per usable result
This article compared 4 popular APIs with an 800-request benchmark.
Good reference: zenrows.com/blog/best-web-sc…
Your favorite tech creator retweeted
A lot of “web-aware agents” are still:
fetch(url) → dump page into model → hope
Production gets messy fast.
JS, bot protection, noisy HTML, wrong page content, extraction formats.
This Mastra + Zenrows setup treats web access like infrastructure.
zenrows.com/blog/web-aware-t…
Your favorite tech creator retweeted
5 code review habits you should steal today for AI-generated PRs:
1. Read the ticket first
2. Review schemas/types before dependent code
3. Separate logic from formatting noise
4. Hunt assumptions, not syntax
5. Give agents a diff-size budget
CodeRabbit breaks it down here: coderabbit.ai/guides/code-re…
Your favorite tech creator retweeted
AI-assisted PRs are 2.6× larger on average in a recent benchmark data.
Once the diff gets huge, review quality drops fast.
A simple rule is to give coding agents a diff budget.
If the task crosses it, split the PR.
See the benchmark here
coderabbit.ai/guides/code-re…
Your favorite tech creator retweeted
And when you're not "conventionally beautiful", you'll see shege. Any bad thing that happens when you're around, it go be like say na your bad aura cause am.
Your favorite tech creator retweeted
Everyone's building AI lead gen agents, but how many are pulling from the same prospect databases?
This tutorial uses OpenAI Agents SDK + Zenrows to discover 78 companies from a live directory, then enrich and score prospects against an ICP.
Code included:
zenrows.com/blog/ai-lead-gen…
Your favorite tech creator retweeted
Persistent coding-agent memory has a nasty edge case:
same error message, different root cause.
Retrieval can find the “right” old memory and still suggest the wrong fix.
I’d love to see this benchmarked adversarially.
Actian’s current Astra implementation: actian.com/blog/developer/bu…
Your favorite tech creator retweeted
Your lead gen agent can get HTTP 200 and still extract zero leads.
Some protected directories return a challenge page instead of company listings.
This Zenrows + OpenAI Agents SDK tutorial shows how to fetch, extract, enrich and score real prospects:
zenrows.com/blog/ai-lead-gen…
Your favorite tech creator retweeted
“If I see any of you at Onitsha Head Bridge celebrating that you maintained a 1950s bridge for the South-East, Holy Ghost will kpaiii you there. This country has no conscience. If you see the way the government is showing this bridge everywhere on the internet, you will think it is a newly constructed bridge, without knowing that it is a bridge built years ago. Imagine someone celebrating his car that he parked for over 50 years after repainting it.”
— Pastor Odumeje bl@sts government over celebration of Onitsha Head Bridge
Your favorite tech creator retweeted
Coding agents repeat investigations they've already solved because the fix stays tied to the old session.
We stored test-confirmed fixes in VectorAI DB so GPT-6 Astra searches them first. Session 2 retrieved the Session 1 fix before touching any code.
Your favorite tech creator retweeted
Hit an out-of-memory error on a 7B model? Drop to a smaller model before you drop below Q4. A 3B model at Q4_K_M usually responds more reliably than a 7B model forced down to Q3 or Q2.
This is how to match quantization to your VRAM.
Your favorite tech creator retweeted
900+ technical writers, devs, and creators, all of us building in the same room.
Inside the Hackmamba Creators Community we run Off the Docs sessions and writing sprints.
You don't have to network. You show up, share your work, make friends, and have fun.
Join us: bki.sh/DdVUpVa
Your favorite tech creator retweeted
An AI agent can spend $1.3M in a month and still have nothing to do with your choice of model.
The model is only one part of what makes an agent work in production.
Your favorite tech creator retweeted
Your AI agent fails.
Do you switch models, rewrite the prompt, or add another tool?
Before doing any of that, ask: what failure are you actually fixing?
Your favorite tech creator retweeted
An AI agent searching for customer ID 48291 needs that exact record, not a semantically similar one.
That's where hybrid search helps.
Databricks explains how vector, keyword, and metadata retrieval fit into agent databases:
databricks.com/blog/database…
