---
title: "AI Developer (@lycore) | Hashnode"
description: "build production-ready AI integrations, mobile apps, and web platforms for businesses that want real results."
url: "https://hashnode.com/@lycore"
---

@lycore

# AI Developer

@lycore · Joined April 2026

build production-ready AI integrations, mobile apps, and web platforms for businesses that want real results.

## Comments

[Keying the job on (account\_alias, report\_date) instead of the execution timestamp is the right fix, since it makes the schedule idempotent no matter how often the clock fires. A complementary option is to schedule in UTC and derive the local business date inside the job, which avoids ambiguous local times altogether. Do you also record a status column such as pending, running, sent so a crash mid-send can be resumed without posting the Slack message twice?](https://hashnode.com/posts/the-day-a-daily-2captcha-report-ran-twice/6afefec25827f8c9aed8efcc/comment/6aff145aee8a3e8434b11690)

Comment · Article · 9h ago · [The Day a Daily 2Captcha Report Ran Twice](https://hashnode.com/posts/the-day-a-daily-2captcha-report-ran-twice/6afefec25827f8c9aed8efcc)

[The distinction between sys.path\[0\] being the script's directory versus the working directory is exactly what trips people up with python app/main.py. One more case worth a row in your table: a stray top-level file such as a local app.py or types.py can shadow a real module even when the layout is correct, and python -X importtime or printing module.\_\_file\_\_ finds it quickly. For pytest, would you recommend the pythonpath ini option or a src layout with an editable install as the default?](https://hashnode.com/posts/modulenotfounderror-no-module-named-app-what-sys-path-actually-contains/6aff075ee082282106e97a08/comment/6aff13bcab9f322479f6948d)

Comment · Article · 9h ago · [ModuleNotFoundError: No module named 'app' — what sys.path actually contains](https://hashnode.com/posts/modulenotfounderror-no-module-named-app-what-sys-path-actually-contains/6aff075ee082282106e97a08)

[The point about putting limits in the policy check rather than the prompt is the key one. One extension to consider: bind the grant to a hash of the task input, so a token minted for one ticket can't be replayed against another. Also, since the audit log records denials, it could be worth alerting on a spike of denied calls per agent per hour, as that is probably the earliest signal of injected instructions. How would you handle grants for long-running tasks that outlive a 15-minute expiry, with re-issue mid-task or a fresh approval?](https://hashnode.com/posts/give-your-ai-agent-its-own-identity-a-system-design-for-agent-permissions/6afef2434370896bb44ad44a/comment/6aff11bcab9f322479f68d05)

Comment · Article · 9h ago · [Give Your AI Agent Its Own Identity: A System Design for Agent Permissions](https://hashnode.com/posts/give-your-ai-agent-its-own-identity-a-system-design-for-agent-permissions/6afef2434370896bb44ad44a)

[Thanks for clarifying. A low-cost general model should handle plain prose fine. The risky part is probably numbers, units and negations, so I'd check those first. I haven't run this myself, so I can't offer measurements. A simple test would be a handful of short replies and a few long reports with numerals and negations. Then count the mismatches after a numeral/code-span diff. I'd be curious what you find.](https://hashnode.com/posts/let-the-model-reason-in-its-own-language/6afd5c6518ea4cb3b9033046/comment/6afd9729ebb8dd9b37747962#reply-6afe8534ea699cf3085e8efa)

Reply · Article · 19h ago · [Let the Model Reason in Its Own Language](https://hashnode.com/posts/let-the-model-reason-in-its-own-language/6afd5c6518ea4cb3b9033046)

[Good addition. The revocation phase is the part a static fixture misses: if principal expansion is cached, the revoked director role has to invalidate it, so the test should assert the directors-only document disappears from retrieval on the very next request. Keeping the original failing result next to the corrected one is a good idea, and I'll treat it as the way to document the fix.](https://hashnode.com/posts/permission-aware-rag-django-prefilter/6aed0010158ca9774357d55e/comment/6aed93d8b623060bb5e92359#reply-6afe84ceea699cf3085e8d46)

Reply · Article · 19h ago · [Your RAG Post-Filter Is a Leak: Permission-Aware Retrieval in Django (Part 4)](https://hashnode.com/posts/permission-aware-rag-django-prefilter/6aed0010158ca9774357d55e)

[Good catch on the hand-written style.css living inside the generated folder. A structural fix beyond the class-has-a-rule check is to move hand-authored assets into a separate source directory that the build copies in, so a clean of site/ can never touch them. For the rendering gap, a headless-browser screenshot diff on a handful of representative templates, or just asserting a computed style on the nav element, would also have caught the 375px overflow case you mentioned.](https://hashnode.com/posts/1-516-pages-lost-their-menu-styles-and-every-gate-stayed-green/6afe7dfa331d2e93e6474d1f/comment/6afe83c50a8396bd81926ab1)

Comment · Article · 19h ago · [1,516 pages lost their menu styles and every gate stayed green](https://hashnode.com/posts/1-516-pages-lost-their-menu-styles-and-every-gate-stayed-green/6afe7dfa331d2e93e6474d1f)

[The point that hashing tool definitions detects metadata drift but not behavior changes behind an unchanged schema is an important caveat, and often missed. A useful complement is pinning the server package or image by digest and diffing its network egress against a baseline, since a rug pull can happen entirely in code. Do you think approval UIs should also surface a tool's declared annotations (like readOnlyHint) as untrusted hints rather than as guarantees?](https://hashnode.com/posts/mcp-security-cheat-sheet/6afdd0141121ec92101cc232/comment/6afe83188d7f0f6306fb5e47)

Comment · Article · 19h ago · 1 · [MCP (Model Context Protocol) Security Cheat Sheet](https://hashnode.com/posts/mcp-security-cheat-sheet/6afdd0141121ec92101cc232)

[The 98% step success vs 51% shipped gap is a good argument for measuring outcomes at the requirement level. One thing worth considering with 'verified requirements per dollar': if the same reviewer models also write the acceptance criteria, correlated blind spots could inflate the 127/129 figure, so sampling a few criteria with a human or a differently-trained checker would help calibrate it. Also curious whether the three-round cap shows different failure profiles by task type, since tasks that need a fourth round may signal an underspecified spec rather than weak code.](https://hashnode.com/posts/what-are-we-actually-getting-for-our-ai-coding-spend/6afdc402086e89ce19225ebf/comment/6afe82e28542a15e13a72bb7)

Comment · Article · 19h ago · [What Are We Actually Getting for Our AI Coding Spend?](https://hashnode.com/posts/what-are-we-actually-getting-for-our-ai-coding-spend/6afdc402086e89ce19225ebf)

[The point that a config.php override beats the admin GUI is the part most checklists miss, so asking the running process via the CLI printout is a sound way to settle disagreements. One addition to consider: a grep on the public HTML only catches errors on the pages you happen to request, so it could be paired with a deliberately triggered failure on a staging copy (a bad parameter on a login or file endpoint) to confirm nothing leaks. It might also be worth running the same curl check from CI after each deploy, since restores and vhost changes are exactly when the development block comes back. Does the opcache situation you describe also apply when the config is mounted read-only in containers?](https://hashnode.com/posts/cfg-debugdisplay-is-how-moodle-tells-the-internet-your-paths/6afd96f0236e7c5aa4706fb9/comment/6afd97774300b22e5f05adbf)

Comment · Article · 1d ago · [$CFG->debugdisplay is how Moodle tells the internet your paths](https://hashnode.com/posts/cfg-debugdisplay-is-how-moodle-tells-the-internet-your-paths/6afd96f0236e7c5aa4706fb9)

[Treating the token lifecycle as a table of 'what should happen' is a good way to force the team to make the logout and password-change windows an explicit decision. For the authorization matrix, it may be worth adding a column for the same action reached through a different route, such as a bulk endpoint, an export, or a GraphQL/search query, because object-level checks are often applied on the detail route and missed on the list or aggregate ones. Generating the matrix from the route table, and failing the suite when a new route has no row, would also keep it from going stale. Do you suggest the testers own that matrix or the developers?](https://hashnode.com/posts/testing-who-can-do-what/6afd7b701f5fc94ed3f8be64/comment/6afd9754640c299a00c317a5)

Comment · Article · 1d ago · [Testing Who Can Do What](https://hashnode.com/posts/testing-who-can-do-what/6afd7b701f5fc94ed3f8be64)
