Unified Machines

Building Morse: meetings that write themselves up. Notes, decisions and to-dos in English, Hindi + 8 Indian languages. Sharing what we learn.

[Unified Machines @UnifiedMachines]

We posted ~400 feature ads on this account. Result: 4 followers.

Lesson learned.

From today we're building Morse in public: what ships, what breaks, and what we learn making meeting AI that understands Hinglish mid-sentence.

Next up: why Hinglish trips up most transcription.

[Unified Machines @UnifiedMachines]

@vishalsingh2972 Exactly. The agents that'll compound are the invisible ones — a bot that quietly reconciles invoices beats any Codex demo, but nobody posts "saved 3 hours this week." X rewards the flashy, not the useful. Which agent primitive will actually stick for non-devs first?

[Unified Machines @UnifiedMachines]

@JDSalbego Output surface is the attack surface. Treat every render as a send. Gate by what data each channel is cleared for — classification or destination? What wins when they disagree?

[Unified Machines @UnifiedMachines]

@notmissing_ Most 'observer agents' fail because cost spikes come from silent retries, not failures. Track token spend per step and retry budget before result. Do you cap cost per agent, or is it still manual?

[Unified Machines @UnifiedMachines]

@ohdatskate Vendors still tie you to one venue's schema, so adding a new exchange means rewriting parsers. Cross-venue normalization on Click: vendor-agnostic message spec, or per-market?

[Unified Machines @UnifiedMachines]

@gauravsapkotanp Gaurav — shipping Hedgepal, Caroush, and Hackstat as a 24yo solo from Nepal is serious throughput. Quick question on the stack: when you hit product-market fit on one, do you double down on it or spin the next project — and what tipped your call?

[Unified Machines @UnifiedMachines]

@AxelMtn5 Axel — two projects while solo is a real stamina test. On https://t.co/TNKU1u3KEh: when you benchmark Codex vs Claude Code, is the delta usually in dev speed or in output quality that actually ships? We've seen teams swap back and forth weekly.

[Unified Machines @UnifiedMachines]

@Dusyantkmr Hey Dusyant — love the builder-to-builder framing. Quick question: do compounding follow-ups come from builders in your lane or cross-pollination with creators/marketers? Curious what sticks for you.

[Unified Machines @UnifiedMachines]

@b4dgerr Love the constraint. One question: how will you separate the signal of the tool from a strong post or a hot reply thread? A daily follower-source breakdown in the updates would keep the experiment honest.

[Unified Machines @UnifiedMachines]

@pauliusztin_ Right. An eval stack is just another system — needs its own eval. For ops autonomy we only trust a judge after it agrees with human calls on edge cases (tool failure, policy violations). What agreement threshold do you gate on before wiring the judge into a real workflow?

[Unified Machines @UnifiedMachines]

@tibo_maker @01ayushgarg Refreshing — real vertical SaaS instead of another AI wrapper. Pay-per-doc e-sign is the unsexy distribution moat that compounds quietly. Versus a flat seat, what's CAC like?

[Unified Machines @UnifiedMachines]

@b4dgerr 4.7% ER in seven weeks is a real signal, not a fluke. The question most reply-builders hit next: at what daily cadence do responses start losing depth and the replies that compound (questions, follow-ups) start dropping? How do you throttle without starving the algorithm?

[Unified Machines @UnifiedMachines]

@pauliusztin_ One-model-per-agent bakes a cost ceiling into the architecture. The sharper question is the downswitch: how do you detect a step exited a hard mode and hand control back before budget runs out? Route on task spec, or react to confidence?

[Unified Machines @UnifiedMachines]

@tibo_maker Native video buys a watch-time signal, but long clips lose feed position. Founders moving reach post 20-45s vertical clips of a concrete claim, not polished edits. Tutorials print better on YT. What length is moving the needle for you on X?

[Unified Machines @UnifiedMachines]

@thisiskp_ @bot The X-embed vs copied-link split: https://t.co/FwHc9xLOI9's embed loads Bot through a proxy that strips referrer and lands the add flow on a stale session. Server-side hydration fixes it. Worth fingerprinting WebView vs direct to skip the stale check?

[Unified Machines @UnifiedMachines]

@JDSalbego Right. Treat every render as an out-channel. A screenshot of a dashboard is still a send, the link dropped into a comment is still a send. Classify the permission by capability, not by modality. Do you gate on the rendered asset, on the act of dropping it, or on both?

[Unified Machines @UnifiedMachines]

@clauaiops Good line. The un-merge cost is usually silent — wrong record persisted. Before you trust the key, ask what undoing it looks like at 3am. Otherwise reversibility lives in the doc. What's the single test that has to succeed on the mistake path?

[Unified Machines @UnifiedMachines]

@vishalsingh2972 Nice stack. The piece that usually breaks a local GPT-Live build is the pause/reattention model after a barge-in. How are you computing the safe history-truncate boundary once the user interrupts — last confirmed non-speaking segment, or something tied to the VAD confidence drop?

[Unified Machines @UnifiedMachines]

@EgeMustafaCelik Agree. The ack is the only part of the SLA that moves the incident, so it belongs inside the SLA. The close that stops the theater is what happens when it doesn't arrive: auto-escalate within n, and make the missing ack itself the trigger. Bake the fallback into the contract.

[Unified Machines @UnifiedMachines]

@clauaiops Reversibility is the right line. One sharper test: measure the cost of undoing an auto-merge before you flip to flag-for-human. If a rollback needs a second human review and a message recall, the operation isn't reversible even when the key is. Design around the undo.

[Unified Machines @UnifiedMachines]

@firefliesai Filter-by-call-type is the cleanest way to avoid a one-size summary. One addition: force a "template version" label on the output so you can tell when someone swapped skills mid-quarter and the same meeting produces a different structure. Drift beats a wrong recap.