Source notes
Smart Rant
How this project works, the decisions behind it and its known limits. Ask quotes and cites these passages when it answers.
Notes reviewed 20 September 2026
Explore the projectProduct
Product
Say what is on your mind. Find the important parts later. Smart Rant makes recording a thought immediate and makes the archive useful later. It is designed around a person’s own moments, people and recurring subjects rather than forcing every recording into a journal template or a coaching conversation.
The problem it solves
A long voice memo is easy to save and hard to revisit. Automatic summaries can be worse when they invent a personal trait, overstate a pattern or lose the evidence behind a sentence. Capturing the moment must also succeed independently of a network request or an expensive model call.
The experience
The iPhone app saves a recording locally first. Normal, Private and Burn modes represent different handling choices. Consented processing can produce searchable transcripts and evidence-backed views across moments. Asking the archive should retrieve the person’s actual recordings and provide receipts instead of presenting an unsupported psychological interpretation.
Availability
Smart Rant is in an open TestFlight beta on iOS, with a web app alongside it. It is not yet an App Store release.
Engineering
Backend and architecture
The native client uses SwiftUI and a durable local archive. The web application uses Next.js and Clerk. The current backend is a Node/TypeScript API and worker over Neon Postgres with pgvector and private object storage; older Supabase material is historical, not the current deployment design. Work is recorded in a durable database queue with token-bound leases and revision/analysis-epoch checks. The pipeline separates transcription, normalization, embedding, extraction, temporal processing, thread refresh and evaluation.
Technology by responsibility
Client: SwiftUI, Next.js, Clerk. Backend: TypeScript, Neon PostgreSQL, pgvector. Infrastructure: Private object storage, Durable job queue, Netlify web hosting.
Data, trust and failure behavior
The worker checks user consent, rant revision and analysis epoch before sensitive work. A stale worker must not publish results into a newer privacy state. Private-mode wording must distinguish local-only capture from explicitly consented transcription. Capture mode is persisted asynchronously and the worker runs inside a between-jobs time budget, so interruption, lease loss and a slow provider stage are the failure cases that matter most.
Save before processing
The capture controller creates a durable take before recording audio and commits the local recording before invoking follow-on work. Recoverable files remain available when finalization fails.
Different privacy modes
Normal, Private and Burn have different processing rules. Private recordings can use cloud transcription only through the separate explicit consent path; that is not permission to include them in normal archive analysis.
Evidence-backed recall
Archive questions use semantic and lexical retrieval, structured evidence and source-grounded quotations. Unsupported or insufficient evidence should produce a limited answer rather than a confident story.
Durable background jobs
Workers claim token-bound jobs, heartbeat while running, classify retryable failures and record outcomes. A stopped process is recoverable through the queue rather than through wishful post-response work.
Decisions
Keeping a recording independent of the cloud
The capture state machine prepares durable storage before the first audio frame and treats save completion separately from processing completion. Trade-off: The extra states and recovery paths are failure cases of their own. A successful UI animation must not be the only evidence that the recording is safely stored.
Making personal claims only with evidence
Extraction, retrieval and generated wording are checked against source material. Recurrence and trait-like statements require more than one anecdote, and the wording guard rejects several forms of unsupported interpretation. Trade-off: Regex and schema checks are conservative heuristics, not a semantic proof. They can reject useful phrasing and miss a paraphrased overreach, so they can be wrong in both directions.
Limits
Known limits
Smart Rant is not on the App Store yet, and billing is still in progress. It does not promise instant background recovery or flawless personal interpretation. It is not a therapist and makes no clinical claim. Public screenshots use synthetic recordings, never the owner's real archive.
Questions Ask answers from here
What makes Smart Rant more than a voice memo app?
It adds a searchable, evidence-backed archive across recordings and subjects while keeping immediate local capture separate from later processing.
Is everything private always processed locally?
No blanket claim is appropriate. The design separates Private mode from a separately consented cloud-transcription path. Normal archive analysis has its own consent and exclusion rules.
What happens when background processing stops?
A durable job ledger, leases and retry classification allow work to be recovered. That does not mean recovery is instantaneous in every deployment.