Research
dispatches.

Thoughtworks built the Tech Radar for organizations. I built one for a single practitioner — automated weekly scans, priority-scored ADOPT candidates, and a memory system that compounds. Here's the architecture, and a real trace from radar signal to integrated tool in under a day.

The complete write-up: twelve interval-arithmetic certificates, a transport proposition that makes them usable inside a parameter cover, a correction to the upper-bound record, and the proofs behind five dead ends.

I set out to improve a lower bound. I ended up rebuilding the certificate underneath it, and finding that the number widely cited as the record upper bound is not a bound at all.

PersonaPlex-MLX asks whether a personal AI interface can feel conversational at human speed while still doing real, inspectable agent work. The project combines local speech infrastructure, cancellable interaction loops, task contracts, risk gates, and evidence-backed worker execution into a gated research program.

Project Prometheus reframes AI education as simulation rather than explanation: a way to explore capability shifts across rules, neural networks, transformers, agents, and emerging multimodal systems.

Harness engineering treats the agent runtime as a first-class system. The article maps failures to concrete levers: instruction design, tool exposure, context hygiene, sub-agent boundaries, memory, and evaluation.

This project treats a personal AI system as infrastructure rather than a chat surface: memory, procedural recall, scheduled context, measurement, and operating principles that compound over time.

Multi-step workflow orchestration (plan→execute, critique→refine) is widely assumed to improve LLM output. I tested this assumption empirically across 6 models. The answer: orchestration helps at 70B+ but hurts at 30B and below. The inflection point lies between 30B and 70B parameters.

Voice assistants should feel like conversations, not command interfaces. Most implementations use a serial pipeline with 3-5 second latencies. Claudia Voice uses two-tier routing that handles 85% of queries at 700ms while preserving depth for complex interactions.

Engram tests whether cognitive memory mechanisms improve long-term AI continuity. In production use, it combines vector retrieval, markdown traces, temporal reasoning, entity tracking, and relevance gates.