Search across signals
Keyword, semantic and graph retrieval run as complementary channels. Reciprocal-rank fusion combines candidates; a MiniLM cross-encoder can rerank them. Time filters apply across the retrieval channels.
Independent product · In development
Doppel
Local AI, grounded in your work.
A native Mac assistant bringing search, speech and source checks together—with specialized models working within a shared memory budget.
RetrieveFind the relevant source
Combine keyword, semantic and graph search, with time constraints where requested.
RankChoose useful passages
A small cross-encoder reranks candidates before they become retrieved context.
InspectKeep the source in reach
Passage links preserve a route from the assistant’s context back to the original material.
Source review follows its own path: original + revision → scope and meaning checks → a reviewable result.
The problem
Finding the words is only the beginning.
A useful answer may depend on a meeting, a note and a document written weeks apart. Similar wording can conceal a different date, scope or level of commitment.
Doppel brings retrieval and interpretation into the same product while keeping source evidence available for inspection. A proposal, a confirmed choice and a report about an earlier event receive different treatment.
Several focused models, one usable workflow.
Check changing support
When a cited original changes, Doppel compares its earlier and current passages. Scope, named details and numeric qualifications stay part of the check. Unavailable comparisons remain visibly incomplete.
Keep meeting meaning distinct
Local processing separates choices, tasks, proposals and reported events. Evidence quotes and timestamps support review; ambiguous model scores can abstain. A model’s interpretation does not itself confirm a record.
Share a desktop memory budget
Model admission accounts for estimated residency and observed process memory. Idle models can be evicted before another loads. Language and vision use MLX; several smaller inference paths use Core ML.
Match the model to the job.
The configured stack combines existing open-weight models with native runtimes, conversion tooling and product-specific checks.
| Job | Implementation |
|---|---|
| Language & vision | LFM2.5-VL-3B · 4-bit MLX checkpoint |
| Speech recognition | Qwen3-ASR-0.6B · Core ML backend, with a separately admitted MLX fallback for eligible failures |
| Speech output | Kokoro-82M · MLX |
| Embeddings | Jina embeddings v2 small · Core ML path configured for CPU and Neural Engine |
| Passage ranking | MS MARCO MiniLM-L6-v2 · Core ML cross-encoder |
| Entity & relation extraction | GLiNER-Relex base · Core ML, with a work-specific label policy |
| Meaning checks | ModernBERT-large-NLI · Core ML, with source and scope gates |
Recovery & continuity
Preserve the work around inference.
The speech path bounds concurrent work, reuses and clears decoder state, and contains supported native exceptions. Failed meeting processing preserves source audio for an explicit retry.
Conversation compaction stores a disposable working summary alongside the canonical transcript. Edits and branch changes invalidate stale summaries. Token-cache reuse requires compatible lineage and matching prefixes.
These are implemented recovery and consistency mechanisms. Their usefulness depends on the complete workload, model revision and hardware.
Evidence status · Implementation review
This case study describes an independent product under development. The repository includes focused regression and conversion checks for the mechanisms described. This article reports no new comparative accuracy, speed or energy benchmark.
Local model inference is one part of the product. Connected services and web search have their own network behavior.
Inside the implementation.
Running several AI models within a desktop memory budget
Residency, compute placement, cache reuse and recovery in an on-device assistant.
Read the noteChecking whether a source still supports a claim
Original passages, revisions, scope checks and bounded comparisons of meaning.
Read the noteBring us the technical obstacle.
Start with an Engineering Sprint: $17,250 USD, two weeks, one agreed obstacle.
See the Engineering Sprint