Independent experiments, engineering demonstrations and selected client projects. Each entry states the problem, the approach, the constraints and what came out of it.
A / CLIENT PROJECTS
Client projects
Sanitized case studies with the problem, constraints, solution and results.
Client case studies are sanitized and published only with the client’s permission. Current engagements remain confidential. When a case study is cleared, it follows the same layout as the work below: problem, approach, constraints and measured outcome.
Our own model testing, quantization, inference and optimization work.
Independent experiment · Vision
Document understanding
Compact vision-language models for extracting structured information from documents.
Problem
Documents arrive as scans and PDFs with layouts that differ from page to page. The useful information sits in fields, tables and headers, not in running text.
Approach
Compare open-weight vision-language models on the same pages. Prompt for structured output, then score field-level extraction against a held-out set.
Constraints
Readable at document resolution. Small enough for a single device. No page images leave the local environment.
Outcome
A working extraction pipeline with a repeatable evaluation set. Model selection, extraction quality and runtime stay tied to the task, not a general leaderboard.
Model selection · Extraction · Evaluation
Independent experiment · Inference
One model, multiple quantizations
Comparing GGUF quantization strategies and their effects on quality, memory and inference speed.
Problem
A single model can be quantized many ways. The smallest file is not automatically the best fit for a task, and perplexity alone can hide a real change in output.
Approach
Quantize one model with llama.cpp across several GGUF formats, with and without an importance matrix. Compare KL divergence, perplexity and file size on the same token sample.
Constraints
Every format must run on the same consumer hardware. The evaluation sample must stay fixed so the comparison is fair.
Outcome
Adding an importance matrix cut KL divergence by 17–31%. Perplexity moved by less than 1.4σ, so it did not show the change. Both numbers come from the same run on the same 25,500 tokens.
Working examples of model integrations, APIs and deployable systems.
System demonstration · Audio
Real-time voice systems
An integrated voice-processing system with speech recognition, language models and speech generation.
Problem
A voice interface needs one continuous loop: hear speech, understand it, decide a response and speak it back without a visible break.
Approach
Connect a speech-to-text model, a language model and a text-to-speech model behind one pipeline. Keep each stage local and measure the latency at the joins.
Constraints
Streaming input, bounded memory and a response fast enough for conversation. All components must run within the target device.
Outcome
A working end-to-end loop from speech input to spoken output, with the latency of each stage measured separately.
STT · LLM · TTS · Deployment
Demonstration · Deployment tooling
Deployment fit estimator
A browser tool that estimates whether a model fits a device before anything is downloaded.
Problem
Model size on a card does not answer whether the model will run on a specific device. Context, KV cache and runtime reserve change the answer.
Approach
Estimate weights, KV cache and a runtime reserve from the model size, weight format, context length and concurrent sessions. Show the planning total against device memory.
Constraints
No server call and no account. The estimate must run in the browser and state its assumptions.
Outcome
A working estimator on its own page, with a memory verdict and a rough tokens-per-second estimate from memory bandwidth.