Hasham Vakani 🇵🇰

Aspiring Full-stack Developer | React Developer ~
Developer of @thenotepd

[Hasham Vakani 🇵🇰 @Hashamdev]

@redp314 @omarchy Speech-to-text on a 1GHz 32-bit chip with 1GB of RAM is wild. What did you have to cut from the model to make it fit?

[Hasham Vakani 🇵🇰 @Hashamdev]

@arena Does Max route once per session, or can it hand off to a different model mid-task? The 67% of code sessions going to Astra reads like a once-per-session call.

[Hasham Vakani 🇵🇰 @Hashamdev]

@karrisaarinen @linear Would be good to see the reopen rate next to this. Fast fixes only count if they stay fixed.

[Hasham Vakani 🇵🇰 @Hashamdev]

@N01ennn @speedrunjaen Finally a leaderboard where the winning move is writing a good CLAUDE.md and going to lunch.

[Hasham Vakani 🇵🇰 @Hashamdev]

@hwchase17 What happens when a thread starts as an easy fix and turns into a refactor two turns later? Escalate on failure, or ride it out on the cheap tier?

[Hasham Vakani 🇵🇰 @Hashamdev]

@supabase @cursor_ai If you wire this up, scope it with project_ref and turn on read_only until you trust it. A chat message that runs on your backend shouldn't get write access to prod on day one.

[Hasham Vakani 🇵🇰 @Hashamdev]

@hwchase17 One way is to score a weekly sample of real runs instead of a fixed test set. Each miss a reviewer catches becomes a new check. The set grows from production, so the agent cannot tune itself to it.

[Hasham Vakani 🇵🇰 @Hashamdev]

Everyone is debating their default coding agent. The honest answer: it doesnt matter until you ship. Pick the one you can put a real guardrail around, because the cleanest demo and the agent you trust with write access are not always the same one. #AIAgents

[Hasham Vakani 🇵🇰 @Hashamdev]

@IntCyberDigest Funny, but there's a dull lesson for the rest of us too. Agent session logs can hold everything you typed and every file the agent opened, so they need the same access rules as any other sensitive data.

[Hasham Vakani 🇵🇰 @Hashamdev]

A prompt test set built from past failures only proves the prompt survives the past. Sample some real inputs every week and score them too.

[Hasham Vakani 🇵🇰 @Hashamdev]

@mardehaym Review is the new bottleneck. Output goes up while the approval queue stays the same size. Cutting handoffs usually helps more than adding another tool.

[Hasham Vakani 🇵🇰 @Hashamdev]

@sergical Licence keys for MCP servers will be a quiet security headache. Where do you keep them so the agent can use the tool but never read the key?

[Hasham Vakani 🇵🇰 @Hashamdev]

@devops_nk What goes in the ops layer first for an agent, tracing and cost limits or evals?

[Hasham Vakani 🇵🇰 @Hashamdev]

@shivam74689 Chunking choices often decide retrieval quality more than the embedding model. Log which chunks came back for each wrong answer. It tells you fast whether the problem is ingestion or retrieval.

[Hasham Vakani 🇵🇰 @Hashamdev]

@danicat83 Nuking works when you have tests or evals that say what working means. Without them you lose behaviour you didn't know you had.

[Hasham Vakani 🇵🇰 @Hashamdev]

@goyalshaliniuk The gap between a demo agent and a production one is mostly retries, limits and what happens when a tool call fails. The LLM call is the easy part.

[Hasham Vakani 🇵🇰 @Hashamdev]

@competentmen Fully agree. It writes the happy path in minutes and leaves you the auth, the edge cases and the 2am alerts.

[Hasham Vakani 🇵🇰 @Hashamdev]

@merge_api For outputs with no single right answer, a rubric scored by a second model works far better than exact match. Braintrust and Langfuse can both run that judge on live traffic, and writing the rubric forces the team to agree on what good looks like.

[Hasham Vakani 🇵🇰 @Hashamdev]

@mikechoi This is the piece I was waiting for. Polling loops are where long agent jobs burn tokens doing nothing. Make the handler idempotent though, events will fire twice sooner or later.

[Hasham Vakani 🇵🇰 @Hashamdev]

@ClaudeDevs Running several at once is what changed how I plan work. How do you handle two sessions touching the same repo, a branch per task or something smarter?