Home
About me
So AI can be safer, I evaluate its performance. One cool way to do this is to figure out the hidden story inside a human expert's mind when they evaluate performance in their field.
Knowledge-dissection projects
- τ-Discernment — What does a user latently require before an agent takes a consequential action? I lift each user's implicit requirement, buried in task prose, into a typed, provenance-grounded object a grader can score (
UserPreflightRequirements) — surfacing failures a task-completion grader misses. github.com/borisdev/tau-discernment - NoBSmed — What makes a medication or supplement recommendation questionable given the evidence, missing evidence, and relevance to the patient?
- HealthBench audit — What makes a "doctor-approved" answer or grading rubric unreliable? I found fabricated citations and 29 possible patient-harm issues in OpenAI's medical-AI benchmark family. github.com/borisdev/nobsmed-healthbench-audit
- HealthBench EBM failure patterns — Where do frontier models fail on evidence-based medicine, and which failures are worth fixing? I ran GPT-5.2 and Claude Opus 4.8 on an EBM-verified slice of HealthBench and clustered the errors into seven canonical recall/precision failure patterns — each with anchored examples, a root-cause hypothesis, and a distinguishing experiment. github.com/borisdev/healthbench-ebm-verified
- Muni-Resilience-Bench — How much does a specific resilience project (wildfire hardening, flood control, seismic retrofit) lower a city's borrowing cost? An open benchmark for the missing, auditable translation from a project's risk reduction to citywide fiscal impact. github.com/borisdev/muni-resilience-bench
- Wolf Games — What makes an AI-generated murder-mystery storyboard coherent across interpolated scenes?
- PhD thesis — What makes an inequality metric capture nuanced gaps among social groups in the same city? I modeled those relationships as weighted edges.
- SimpleLegal — What makes "Called the State Senator" lawyer-level work rather than an administrative task? The classification depended on the relevance of the output, not the wording of the task.
Selected technical impact
- Audited OpenAI's HealthBench medical-AI benchmark family — surfaced fabricated citations and 29 possible patient-harm issues: github.com/borisdev/nobsmed-healthbench-audit
- Built an AI agent-evaluation framework for Sindri.ai using Temporal.
- Helped relaunch a stalled legal-billing AI feature by eliciting lawyer expertise and redesigning the rubric for 11 billing flags.
- Built story-scene prediction for a gaming startup led by a Law & Order producer.
- Contributed an experimental causal-graph agent to LangChain: langchain-ai/langchain#6255
- Built backend systems for factory analytics, people analytics, and Tableau geospatial services.
- Embedded subjective concerns into statistical analysis of income inequality in my social-science PhD: escholarship.org/content/qt8br7d5df
Non-tech fun points
- Climbed Cotopaxi (about 19,300 ft).
- Survived bodyboarding Mexpipe.
- Taught geospatial data to students in Medellín, Colombia.
- Taught kids snowboarding.
- Managed an international restaurant team.
- Counseled severely emotionally disturbed children.
Elsewhere
- Nobsmed blog — work-in-progress notes on clinical-evidence retrieval and personalized trial matching:
- evidence-to-person-eval — open benchmark for whether Medical AI applies clinical-study findings to heterogeneous people without overgeneralizing
Writing
Essays on the patterns I've seen recur. Some are primarily a way to organize my own learning; I'm not an expert in everything I write about — especially the compiler-design piece.
| # | Title | Topic |
|---|---|---|
| 1 | Beyond RAG: How Chomsky's I-Language and Compiler Design Converge on Knowledge Graphs | LLVM-style IR, Chomsky's I-language, BFO ontology, grammar-first design |
| 2 | What Is Knowledge Engineering, Really? | A working definition built around elicitation, evaluation, and 0→1 modeling in messy domains |
| 3 | Fine-Tuning LLMs Will Restructure Your Data Science Team | How fine-tuning replaces annotation pipelines and the NN-optimization role; the new "fine-tuning analyst" |
| 4 | Why Domain-Specific Language AI Features Fail | The customer-discovery process for niche language AI, and why a Lean Startup approach is required |
| 5 | Language AI Evaluation 101: Know Your User | Why simplistic Ground Truth produces misleading accuracy metrics; cognitive empathy as the iteration loop |
| 6 | Hyper-Local Community Funding: A DAO Alternative to CDFIs | Local digital tokens and DAOs as a delivery mechanism for under-served-neighborhood capital |
| 7 | Inequality with a Spatial View | A note from my 2014 dissertation: the same income data can read as inequality going down or up depending on whether you keep the spatial structure. A spatial view is a graph |
| 8 | CV: Knowledge Engineering in Messy Domains | The IR-compile pattern across clinical trials, legal billing, maritime construction, narrative gaming, and geographic inequality |