HNHacker News
TopNewBestAskShowJobs

Sal4906

1 karma · joined July 23, 2026

submissionscomments
Sal4906··on Ask HN: Who wants to be hired? (September 2026)
Location: London, UK

Remote: Yes, remote only

Willing to relocate: No

Technologies: Python, FastAPI, MCP, Rust, SQLite, TypeScript, Oracle Cloud ARM, Cloudflare

Resume/CV: https://github.com/visione4906

Email: crease.tm@outlook.com

Pick a sector at https://demo.consentleads.uk and message it the way a customer would. Nothing is scripted. That is a production LLM system I built and run on my own: free-tier ARM box, Cloudflare tunnel, scheduled jobs keeping it up.

The part worth asking me about is the suppression gate. One opt-out anywhere suppresses that address everywhere, across every product and every client, and the check runs before every marketing send. It also fails open. If the suppression file will not parse, the lookup returns false and the message goes out. I claimed the opposite in the August thread and corrected it there (https://news.ycombinator.com/item?id=49158009). That repo is private, so treat all of it as a claim rather than proof.

Four things that are public:

agentic-rag (https://github.com/visione4906/agentic-rag). Retrieval as a tool the model decides to call, not a pipeline stage. Hybrid BM25 and dense with RRF, behind an MCP server I wrote by hand, JSON-RPC over stdio. The eval scores refusal as well as recall: 8 of 10 answerable questions at k=5, plus 3 unanswerable ones it refuses rather than answers. The first version scored 9 of 10, on a set that included two questions written from the indexed document's own sections. Replacing them cost a question, and the README keeps the drop in, because an eval written from its own corpus flatters the retriever. Thirteen questions finds a failure mode. It does not rank anything.

demo-e2e (https://github.com/visione4906/demo-e2e). Smoke tests that drive the demo above in a real headless browser, because a health check can read green while the page is dead. A test that cannot fail is not a test, so they were run first against a target with no chat surface to confirm the assertions bite.

durable-media-pipeline (https://github.com/visione4906/durable-media-pipeline). SQLite state machine, crash-safe resume, an idempotent re-run path on the extraction stage. Six scheduled tasks have run it since May 2026. That is what unattended means here.

brain-engine (https://github.com/visione4906/brain-engine). A case-based decision engine in Rust behind a second MCP server, eight tools over stdio. MIT, as is agentic-rag.

Looking for AI engineering or AI automation work, remote. To be straight about it, I start a full-time degree later this month, so I am after part-time or flexible remote work rather than a standard full-time role. That is what the two answers at the top mean.

Sal4906··on Ask HN: Who wants to be hired? (August 2026)
Location: London, UK

Remote: Yes, remote only

Willing to relocate: No

Technologies: Python, FastAPI, Anthropic Claude, TypeScript, Next.js, SQLite, Solidity, Oracle Cloud ARM, Cloudflare

Resume/CV: https://github.com/visione4906

Email: crease.tm@outlook.com

I build LLM systems that run in production. The clearest example is live: pick a sector at https://demo.consentleads.uk and message it the way a customer would. Nothing is scripted. It runs on a free-tier ARM box behind a Cloudflare tunnel, is multi-tenant and has users, and nine scheduled jobs keep it going unattended.

If opt-out detection misses a phrase, the system keeps emailing someone who asked it to stop, and nothing crashes or alerts. The suppression gate fails open, and the eval fixtures cover that path and reply triage.

[Corrected 28 Aug 2026. This originally said the gate fails closed. It does not. If the suppression file fails to read, the code logs the error and returns whatever it had already parsed. If a single line is malformed, it skips that line. Either way the lookup returns false and the send proceeds. The gate is real and runs before every send. The label was wrong.]

I also wrote claimcheck (MIT, https://github.com/visione4906/ctaio-claimcheck): give it a document and a codebase and a second agent has to prove every claim against the source with file and line citations, or the build fails. I built it after an outside review of my own writing turned up claims the code did not support.

Looking for AI engineering or AI automation work, remote.