Insights
Thinking Out Loud
Essays on biotech R&D, AI architecture, and the seams where deep technical work meets executive strategy.
-
Biotech
Single-Cell Foundation Models: Three Adoption Lessons
Roughly 73 percent of new spatial transcriptomics projects reach for a single-cell foundation model. Most hit the same three walls (preprocessing, hardware, regulatory) and discover the model was never the moat.
-
Business Development
The Federal Proposal Section No One Edits Twice
Action items, clarifying questions, and IP novelty get drafted once and never reviewed. That is where roughly 30 percent of proposal losses originate, and where the team's confusion about scope, dependencies, and IP cleanly leaks through.
-
Biotech
When Your Fermentation Data Talks Back: AI-Augmented Bioprocess Development
AI and ML are shifting bioprocess development from trial-and-error to data-driven optimization by giving scientists pattern recognition across thousands of fermentation runs that no human can replicate manually.
-
Biotech
The Synthetic Biology Stack: Why Biotech Needs Software Engineering Principles
Biotech companies that adopt software engineering practices (version control for strains, CI/CD for genetic constructs, automated testing for phenotypes) will outcompete those that don't.
-
Data Science
Validating Synthetic Genomic Data: The Missing Quality Layer
As synthetic genomic data becomes critical for ML training and privacy-preserving research, we need validation frameworks that measure fidelity, utility, and privacy, not just statistical similarity.
-
R&D Strategy
Innovation Driven by Ability, Not Self-Awareness
Capability expansion precedes need articulation. The most consequential technologies in history weren't solutions to articulated problems.
-
Innovation
Starting With What's Possible: Why R&D Should Lead Innovation
The R&D-First Bullseye Model reverses traditional market-led approaches by starting with technical possibility and working outward toward market validation.
-
Business Development
The Gold Mine in Your Lost Proposals
Your organization's lost proposals and unpursued opportunities aren't failures. They're an asset library waiting to be indexed.
-
AI + Tools
Memory Is a Snapshot, Not a Source of Truth
Your AI's memory is a snapshot of what was true at some past moment. The code, the file paths, the function names, the flags all move. The memory does not. Verifying a recalled fact before acting on it is the cheapest discipline in the toolkit and the one that gets skipped most often.
-
AI + Tools
Web-Search Seed, Local Synthesis
If your LLM needs current web data but you can't give it internet access, the answer is not to find an LLM with internet access. It is to separate retrieval from synthesis. The retrieval phase has internet. The synthesis phase has the data. Neither phase has both.
-
AI Engineering
Cloud LLMs Have Failure Modes Benchmarks Don't See
Average-case benchmarks tell you what the model does on a typical day. They do not tell you about the failure mode that fires on one dispatch in four and tanks your output. A 4,000-token loop of 'incumbency procuring obtaining acquiring' does not show up in a benchmark score.
-
AI Engineering
The Single-Reviewer Trap
One reviewer LLM gives you the model's loudest opinion, not the truth. A panel of three with different framings surfaces what one alone flattens. The cost of three parallel reviewers is small. The cost of missing the failure is whatever it costs when the artifact lands live.
-
AI Engineering
Recon Before Build: The Cheapest Insurance in Agent Dispatch
Small verification dispatches before a real build cost about 3 percent of your context window. They prevent 5 to 10 percent burns on bad-premise builds. Across a session, that compounds into the difference between finishing and running out of room.
-
AI Engineering
Why I Run Three LLM Backends and Mostly Use the Cheapest One
Most teams pick one LLM backend and stick with it. That is backwards. The interesting number is not which model is best. It is what fraction of your actual work needs that model. In my practice, usually 5 to 15 percent.
-
AI Engineering
The Knowledge Base Nobody Reads (Including You)
Your company has a 70MB knowledge base aggregated across 14 collections. It is consulted maybe five percent of the time when relevant prior work exists. That is an audit trail, not a knowledge base.
-
AI + Tools
Bubblewrap for LLMs: Why YOLO Defaults Lose
Every AI coding agent ships with full filesystem access, full network reach, and permission to read every environment variable your shell knows. That works at hackathon speed and fails the moment the data matters.
-
AI Engineering
Five Models, One Spec: Notes from Routing the Same Work to Different LLMs
When you dispatch the same tightly-specified task to GPT-5, GLM-5.1, GLM-4.6, and DeepSeek-V3 in parallel, the failure modes are model-specific in ways that matter. A working set of rules for choosing the right model per task.
-
AI + Tools
The Syntax Tax
Enterprise and scientific software has a long tail of capability most teams never reach because it is locked behind syntax. AI coding assistants change where in the workflow that translation happens.
-
AI Architecture
Parallel Subagent Dispatch and the One Failure Mode That Hides
The bottleneck in parallel LLM dispatch is not parallelism mechanics. It is independence detection. Most subtasks that look independent are coupled through the filesystem, and the failure mode that lets multiple agents corrupt each other does not announce itself.
-
AI Architecture
Three-Way Dispatch: Write, Review, Compose
Most LLM routing discussions are about cost tier. The discussions that matter are about cognitive role. Writing code, reviewing code, and composing the next prompt are three different jobs and benefit from three different models.
-
AI Architecture
The Hidden Tax on AI Agent Swarms
Decomposing work across multiple AI agents often costs more than a single well-designed agent. Here is why that happens and what the numbers actually look like.
-
AI Architecture
The Git Repo as the AI Workflow System Boundary
A git repository is the natural and complete system boundary for AI-driven document workflows. Everything the agent needs already lives in or can be referenced from a repo.
-
AI Architecture
What Is the Smallest Unit of Work in Agent Decomposition?
When decomposing tasks for AI agents, there's a granularity floor: a point below which further decomposition creates more overhead than value.
-
AI Architecture
Knowing That You Know: External Memory Architecture for AI
Most AI memory architectures have two states: knowing and not knowing. A third state, knowing that you know, lets systems scale beyond their context window without loading everything at once.
-
AI Architecture
The Context Window is a Viewport, Not a Bucket
Knowledge map traversal (navigating structured graphs on-demand instead of loading entire databases) is the superior pattern regardless of context window size.