AuraLinter
Will It Sound Good
When You Post It?
AuraLinter is a recording-quality coach for creators. Record or import audio, get evidence-backed verdicts on intelligibility, clipping, noise, and echo — then hear and export the corrected version. Plus: ask DSP questions and get compiler-verified C++ kernels from a multi-agent LangGraph backend.
How It Works
From microphone to compiler-verified C++ kernel — an agentic DSP pipeline on your iPhone.
Engineered for DSP Intelligence
A SwiftUI frontend backed by a FastAPI + LangGraph agentic orchestration engine — with compiler-verified code generation and Production RAG.
Recording Quality Coach
Record or import audio, select your intent (Video, Podcast, Voice), and get evidence-backed verdicts: speech coverage, loudness, clipping, noise, room echo, and intelligibility — each with confidence scoring and actionable fixes.
Evidence, Not Vanity Scores
Every finding comes with evidence: framewise envelope waveforms, speech/silence intervals, clipping markers, and measured values. cant_tell verdicts are honest when the model lacks confidence — no fake metrics.
A/B Preview & Export
Toggle between original and corrected audio with hold-to-hear-original. Export to Video MP4, Audio M4A, or WAV — all with loudness normalization applied to your target level.
Agentic DSP Engine
Switch to the Ask tab for multi-agent DSP code generation. A LangGraph state machine with Coder → Critique loop generates C++ kernels, verifies them with clang++ -fsyntax-only, and streams the process live via SSE from a FastAPI backend. Production RAG over the DSP canon with hybrid pgvector search.
Core Capabilities
Each capability is a complete surface: capture, query, generate, and verify — all orchestrated by agentic AI.
Onboarding
A 3-page guided tour introducing the creator recording-quality loop: check, verdict, fix. Honest expectations about what the app can and can't measure.
Check Tab
The main entry point. Record or import audio, select an intent (Video, Podcast, Voice), add a reference track, and submit for AI-powered quality analysis.
Live Processing
Watch the LangGraph agent work: router classification, loudness measurement, EQ testing, compressor evaluation — each step streamed live as it happens.
Quality Results
Evidence-backed findings: speech coverage, loudness, clipping, noise, echo, and intelligibility. Each card shows the verdict, confidence, evidence, and actionable fix.
Editor & Export
A/B preview: toggle between original and corrected audio with hold-to-hear-original. Export to Video MP4, Audio M4A, or WAV with loudness normalization applied.
Projects
Local SwiftData persistence plus server-side run history. Reopen any past check, revisit results, or continue an editor session exactly where you left off.
Ask (Agentic Q&A)
Ask DSP questions, attach audio for context, and get compiler-verified C++ kernels. Full LangGraph multi-agent pipeline with clang++ critique loop.
Settings
Configure backend endpoint and API key, manage local storage per-project, view server run history, and access support resources. Self-host or connect to managed.
Built For
Whether you're shipping an audio product, publishing DSP research, or learning the fundamentals.
Audio Developer
- Generate custom C++ DSP kernels with compiler verification
- Query the DSP canon for implementation patterns
- Iterative critique loop with clang++ syntax checking
- Export verified kernels directly from the app
Researcher
- Production RAG over DSP textbooks and papers
- Hybrid pgvector search with source citations
- Structured LLM routing for complex queries
- LangGraph state inspection for reproducibility
Student
- Learn DSP by seeing AI generate and verify code
- Watch agentic reasoning unfold in real-time
- Query textbooks with natural language questions
- Understand C++ filter design through interactive iteration
Technical Specifications
Under the hood of the agentic DSP orchestration stack.
| Frontend | SwiftUI (iOS 17.0+) • AVAudioEngine capture • URLSession SSE streaming |
| Backend | FastAPI + LangGraph StateGraph with Postgres checkpointer |
| LLM Provider | Amazon Bedrock (switchable to Anthropic direct) • Haiku router • Opus coder/judge |
| Vector Store | pgvector (HNSW dense) + keyword ts_rank • RRF merge • Voyage AI embeddings (1024-dim) |
| Code Verification | Sandboxed clang++ -fsyntax-only • bounded critique loop • graceful failure states |
| Audio Analysis | librosa (STFT, MFCC, delta/delta-delta, spectral descriptors, pitch) |
| Observability | Arize Phoenix OTel tracing • structured logging • token-usage tracking |
| Auth & Security | API-key auth middleware • configurable rate limiting • CORS by explicit origin list |
| Deployment | Docker • Fly.io • docker-compose local dev |
| Hardware | iOS 17.0+ • iPhone / iPad |
Frequently Asked Questions
Real answers about agentic DSP orchestration.
clang++ -fsyntax-only, and a Judge decides whether to accept or retry. The entire process streams live to your iPhone via SSE.
LLM_PROVIDER environment variable — direct Anthropic API is one env var away. Model tiering (router/coder/answer/judge) is fully configurable.
ts_rank search. Results are merged via Reciprocal Rank Fusion (RRF) for the best of both worlds. Every RAG answer includes numbered source citations from the ingested DSP corpus, so you can verify exactly where the information came from.
docker compose up -d db && uv run uvicorn app.main:app --reload, or deploy to Fly.io, AWS, or any Docker-capable platform. Configure your iOS app's Settings to point at your own backend endpoint with your own API key structure.
