Wahaj Ali

Hey @X algorithm Show this to people who are actually building: • AI agents • Production systems • LangGraph / FastAPI • Real AI employees Drop what you’re working on 👇 Let’s connect.

Testing Velix against different data structures exposed cases the existing approach could not handle. I designed an improvement workflow and tested it end to end in development. That validates the workflow in those tested cases. It does not establish support for every unfamiliar structure. My takeaway: revisit the failing case after the change, rather than only demonstrating examples that already work. Velix is a ProveinTech product in development/testing.

Alhamdulillah! Passed with 733/1000! My journey preparing for the Claude Certified Associate Foundations certification: 1st attempt: 283 (unprepared) 2nd: 450 3rd: 683 4th: 600 5th: 733 PASSED! Covered AI fundamentals, prompting, agentic workflows, Skills, Connectors, evaluation, workflow design, governance, security, and human oversight. Five attempts. Countless lessons. One milestone. Never give up. Alhamdulillah! 🤍 #AgenticAI #AIEngineering

Building Velix has reinforced an integration lesson: The same business operation can sit on very different database structures. I chose to work toward adapting the application to existing systems, rather than asking businesses to restructure their databases around it. Testing exposed unsupported cases. One working example does not establish broad compatibility. Velix is a ProveinTech product in development/testing. Compatibility is still being evaluated.

A CMS publish succeeded, but refreshing the public site briefly showed older content while its configuration API request was pending. I chose server-side rendering for the initial page so configuration could be fetched before the page was rendered. That addressed the old-content flash I observed. Trade-off: configuration fetching becomes part of the initial response path. This was a rendering fix, not a universal cache-freshness solution.

A CMS snapshot bug I caught during verification: Frontend logic inside a closure decided whether an update should create a snapshot. Sometimes it did. Sometimes it did not. I removed that decision from the frontend, moved it to the backend and verified the corrected snapshot behavior. The issue was not closures themselves. It was letting frontend logic decide whether required version history was preserved. I reviewed and fixed the existing implementation; I did not build the original feature.

Velix's intended fulfillment notification rule: Owner marks an order fulfilled through a natural-language prompt → email the order's category manager. Category manager does it → email the owner. The initiating role determines the recipient. End-to-end testing is pending. This describes the workflow design, not verified email delivery. Next check: both role paths, the resulting order state and receipt by the intended person. Velix is a ProveinTech product in development/testing.

For Velix, I defined conversation memory that stores queries and responses, scopes them by tenant, and retrieves relevant exchanges for later requests. I used an AI assistant to generate the implementation code. Recall, missing-memory behavior and cross-tenant isolation are still untested. One boundary matters: recalling an earlier stock answer is not verifying historical inventory. Next checks: relevant recall, no-match behavior and cross-tenant access attempts. Velix is a ProveinTech product in development/testing.

Two things I added to my personal News Agent: 1. Workflow steps in the UI. 2. Response streaming across the backend and frontend. I revisited the demo: steps appeared first, then the response arrived chunk by chunk. Steps communicate processing stages. Streaming progressively displays the answer. I have not measured latency gains or user comprehension. A responsive-looking interface also does not establish answer accuracy.

Been looking at how the X algorithm actually works. Replies and shares seem to count for a lot more than likes. At 434 followers and wondering if 1k in a day is realistic, or if it’s mostly steady daily work. What actually moved the needle for you?

I asked ChatGPT to show how I've treated it over time. Here's what it came up with. 👇

A screenshot should not require the user to retype the claim. I added OCR input to my personal News Agent: Image → extracted text → research workflow → LLM assessment. OCR handles extraction. It does not verify the claim. I have not benchmarked extraction accuracy across screenshot types. A misread input is a different failure from a poor assessment. The useful design question: what work can the application take off the user's hands before the main workflow even starts?

My News Agent started with the Guardian API as its research tool. I added a fact-check API and Wikipedia because the questions I wanted it to handle needed evidence beyond one source. The boundary: Tools retrieve information. The LLM assesses the retrieved evidence. More sources do not guarantee a correct judgment. I did not establish deterministic source weighting or calibrated confidence scoring. This is an independent personal prototype, not a production verification service. What tells you a research workflow needs another source?

We build AI agents, automation systems, and websites. Another agency just emailed us offering AI integrations, automation, and website development. 😭 B2B outreach has officially entered an infinite loop.

Noticed an interesting navigation behavior in the X mobile app (Subscription → Analytics). After switching through the top tabs (Overview → Audience → Content → Spaces) and pressing back, it immediately exits Analytics and returns to the main Subscription/Plus screen instead of stepping back through the visited tabs. This typically occurs when tab switches are treated as internal state of a single destination rather than being pushed onto the navigation back stack. In standard mobile navigation (Android Navigation Component / iOS UINavigationController), we normally expect LIFO behavior if each tab is modeled as a distinct destination. Sharing the recording in case the team wants to review the current navigation approach. @X @Support Happy to share more details if useful.

I found an OTP state bug in Velix: successful verification was not setting the database's verified flag. An unverified user could also continue into setup. I fixed the flag update and added a backend login check: Unverified account → no access token → new OTP → frontend verification screen. This is the implemented login behavior, not a claim of complete authentication security testing. The lesson: check persistence and enforcement, not just the OTP screen. Velix is a ProveinTech product in development/testing.

Just crossed 400 followers. Thank you to everyone who’s replied, connected, or even just scrolled past while I’m figuring this out in public. I’m building agentic AI stuff day-to-day RAG, agents, integrations and shipping Velix with the team at ProveinTech. Most days it’s debugging state mismatches, trimming unnecessary agent steps, and trying to make the 10% human / 80% AI / 10% human loop actually work. If you’re also building with agents, RAG, or just trying to grow on here while shipping real things drop what you’re working on (or what’s currently broken). I’d love to connect with more people in the trenches. What’s one thing you’re building or learning right now?

Backend: 401 Unauthorized. One frontend hook: unauthorized. Another hook: isLoggedIn = true. In Velix, that disagreement became an infinite redirect loop. I traced and fixed the state mismatch. I do not remember the exact code change well enough to share a reproducible fix. The debugging lesson: inspect how session state changes across hooks, not just what each hook does individually. Separate bug from the onboarding screen flash. Velix is a ProveinTech product in development/testing.

Onboarding was incomplete, but Velix briefly showed the dashboard. A few seconds later, it redirected to onboarding. I moved the initial auth/onboarding checks and routing decision to the server side. I no longer observed that flash in the flow I checked. The final destination was correct. The initial screen was the problem. This was a UI routing fix, not evidence of an authorization bypass. Velix is a ProveinTech product in development/testing.

I reviewed a plan to turn an existing application into a multi-tenant SaaS. The proposed data model had no tenant entity or table. I corrected the architecture document. The project did not proceed to implementation. My first review question now: “Where does the tenant actually exist in this model?” Then I can examine how users, business data and permissions relate to it. An explicit tenant table was the correction in this design. It is not, by itself, proof of tenant isolation.

Small requests were triggering unnecessary agent steps in a Claude-based development workflow. I traced the token usage. Subagent overhead and repeated loading of irrelevant history stood out. I added: • Markdown indexing and links • Selective-reading rules in CLAUDE.md • A script that prepared relevant context summaries for defined small-task intents Larger tasks retained the broader agent workflow. Across my tested before/after scenarios, total tokens dropped: 63% for small tasks 19% for larger tasks Scenario-specific results, not a general benchmark. Output quality was not formally measured. What context does your workflow load unnecessarily?