Technical Partner

1. Scope of the work

Model behavior is only part of the problem. Memory, permissions, bad inputs, latency, cost, and recovery often decide whether a product can be operated. I review those constraints before a team commits to an architecture.

I do not offer customer discovery, market sizing, pricing, sales, or product-market fit work.

2. Independent research

Three full papers are available here. The first defines a tenant-isolation protocol for conversational memory. The second audits a preserved 46-case run of a restaurant agent. The third documents an extraction pipeline, reproduces its 569 software tests, and states the benchmark that still has to be run. These are independent working papers. None has been peer reviewed.

1. WP-01 · Working paper · 2026

   Tenant-Scoped Long-Term Memory in Conversational Agents: Architecture, Threat Model, and an Executable Evaluation Protocol

   Can a multi-tenant conversational agent recover durable user facts without exposing facts from another tenant, and can that claim be tested with an inspectable protocol?

   Evidence status: Protocol and implementation inspected; no public live-run artifact

   Read paper →

2. TR-01 · Technical report · 2026

   Failure-Focused LLM-as-a-Judge Evaluation for a Transactional Conversational Agent: A 46-Case Repository Study

   Can a category-based model judge expose actionable failure clusters in a multi-turn transactional agent, and what must be added before its scores are reliable evidence?

   Evidence status: Saved 46-case run inspected; provenance gaps and test-design limits documented

   Read paper →

3. RP-01 · Research protocol · 2026

   Provenance-Constrained Multi-Stage LLM Extraction from Heterogeneous Product Catalogs: System Design and Benchmark Protocol

   Does separating document understanding, source-row reconstruction, and product normalization improve reliability over a single-pass extraction prompt?

   Evidence status: Architecture and 569 software tests reproduced; extraction benchmark still unexecuted

   Read paper →

Open the research index →

3. Ways to work together

Bring one decision that could change the architecture, cost, or feasibility of the product. I review the material, agree with you on what would count as an answer, and return a recommendation you can use. The USD 120 session fee is credited toward a sprint if I accept it.

4. How I reach a recommendation

First we write the decision in plain language and name the result that would change it. I separate facts from assumptions, then choose a test small enough to run but strong enough to rule an option in or out. The final brief states the recommendation and where it stops being reliable.

5. Boundaries and fit

• One engagement addresses one primary technical question. One sprint can be active at a time.
• The work can assess technical feasibility and implementation risk; it does not validate demand, pricing, distribution, sales, or product-market fit.
• No engagement guarantees funding, revenue, adoption, production success, or any commercial result.
• Complete MVP delivery, open-ended development, and indefinite support are outside scope.
• Logistics, last-mile, routing, fleet, dispatch, delivery management, delivery tracking, and adjacent competitive work are excluded.