Biology and the next experiment
On fermentation data, synthetic genomes and making sense of model output.
Essays
- Validating Synthetic Genomic Data: The Missing Quality Layer
- When Your Fermentation Data Talks Back: AI-Augmented Bioprocess Development
Insights
Essays on AI workflow engineering, automation, and data science, and the deep-technical work that proves the method transfers across domains.
Selected essays
On fermentation data, synthetic genomes and making sense of model output.
Essays
How to organize project records and trace the numbers in AI-assisted work.
Essays
Articles on process data and the research decisions that follow from it.
Essays
Where modest tools can save repeat effort and help a team reuse past work.
Essays
Archive
AI Engineering
A team can get value from AI chat without changing a recurring business process. The gap is often the work of defining inputs, checks and a person responsible for the result.
Data Science
A traceable data-assembly workbench can help prepare inputs for engineering modeling tools. The value must be measured in a pilot, not assumed from a demonstration.
Data Science
An engineering draft is easier to review when important values trace to a source or are explicitly flagged. That discipline helps distinguish a demonstration from something a professional can evaluate.
AI + Tools
For science and engineering teams handling sensitive information, data boundaries can matter as much as model quality. A written policy is useful only if the actual workflow enforces it.
AI Architecture
A good first AI workflow is frequent enough to matter and simple enough to verify on representative cases. Risk, ownership and maintenance costs still influence whether it is worth building.
AI Engineering
A green test suite tells you the model's own logic holds. It says nothing about whether the output survives contact with your real workflow. Here is the verification ladder I use before anything ships.
Biotech
Machine learning can help bioprocess teams recognize deviations and test hypotheses across historical runs, provided the models are validated for their intended decisions.
Innovation
AI demand has real grid, water and climate consequences. Energy-per-query figures alone cannot tell us whether its benefits justify those costs or what infrastructure should be built.
AI Engineering
In a blind legal-domain evaluation, a four-model ensemble contradicted source material less often than human-written interpretations, but produced more uncited references. The trade-off matters.
AI Engineering
Data collection and reconciliation are plausible targets for bounded automation around engineering work. The effort saved and the accuracy of the output still need to be measured.
AI Engineering
Unanet's 2026 AEC survey reports widespread AI adoption but far less confidence in source data. Those are different measures,and an important design problem.
AI + Tools
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 Engineering
A polished AI-assisted artifact may be easy to produce without being ready to use. The decisions, rejected assumptions and tests behind the result tell you far more than its appearance.
Biotech
An internal AI tool is not automatically covered by FDA's draft AI credibility framework. Its intended use, impact on regulated work, and existing quality obligations determine the review.
Federal Funding
Federal proposal teams often rebuild the same operating structure from one cycle to the next. More writers add capacity, but they do not give the process a memory.
Business Development
A lost proposal may contain reusable technical explanations or prior decisions. Reuse requires permission, current facts and a record of what was never validated.
R&D Strategy
Capability expansion precedes need articulation. The most consequential technologies in history weren't solutions to articulated problems.
AI Engineering
A knowledge base that gets updated after every project but rarely consulted before the next one is an audit trail, not a working memory.
Innovation
Karpathy says we'll never detect AI in homework. He's right, and it exposes a problem older than AI: the proxy we'd been grading on was always thin, and we kept using it because the expensive version of the test was unaffordable.
Business Development
Action ownership, unresolved assumptions, and the claim of novelty can expose weaknesses in an R&D proposal even when its technical narrative is polished.
AI Engineering
A short verification pass before a costly build can catch a stale function name, a wrong file path, or a bad architectural assumption before the implementation starts.
Innovation
The R&D-First Bullseye Model reverses traditional market-led approaches by starting with technical possibility and working outward toward market validation.
AI + Tools
Samsung's reported source-code uploads exposed a practical problem: staff need to know which data may enter which AI tools. An unsupported 42% leak statistic is not needed to see the risk.
R&D Strategy
A practical approach to scientific technical writing, from choosing the right level of detail to checking claims and revising with readers.
Biotech
Preprocessing, compute, and biological validation, not arbitrary cell-count thresholds,determine whether a single-cell foundation model adds value.
AI Engineering
Three AI reviewers given different briefs found different problems in one strategy draft. That's a useful demonstration of framing, not proof that three reviewers are always better than one.
Data Science
Before any engineering judgment happens, a municipal drainage study runs a long assembly line of manual data work. Here are the six places the hours actually go.
AI Engineering
NCEES calls for independent checks of AI-generated material. Its position statement is guidance, not state law,and insurance coverage depends on the actual policy.
Biotech
Strain lineage records, reproducible construct checks and staged phenotype tests can make biotech development easier to audit. Biology still needs experiments, not software metaphors alone.
Data Science
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.
AI + Tools
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.
Federal Funding
Five criterion-linked checks for pressure-testing whether SBIR Specific Aims make the innovation, approach, evidence, risk, and team logic easy to evaluate.
R&D Strategy
What goes into a techno-economic analysis, what its results can tell you, and whether a TEA will help your next process decision.
Engineering notes
Implementation-internal essays on agent dispatch, context windows, model backends, and memory systems, the tooling this practice runs on rather than the client problem it solves.
AI + Tools
An AI coding agent's access depends on how it is configured. Sandboxing narrows its blast radius, but boundaries must be verified rather than assumed.
AI Engineering
Average-case benchmarks do not tell you which failure modes appear on the long, structured work you actually need to ship. Test models on your workload and inspect the tails.
AI Architecture
For questions involving explicit relationships, on-demand graph traversal can complement semantic search and spare context space. It carries its own indexing and query costs.
AI Architecture
A git repository can be a durable coordination boundary for AI-driven document workflows, but it is not a security sandbox or a complete record of agent activity.
AI Architecture
Splitting work across more AI agents adds context, prompt, review, and handoff costs. Measure those costs before assuming that more agents mean more useful parallelism.
AI Architecture
Metadata about information stored elsewhere can guide on-demand retrieval without loading an entire collection into context. It requires reliable indexing and still has failure modes.
AI + Tools
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 Engineering
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 Architecture
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
When decomposing tasks for AI agents, there's a granularity floor: a point below which further decomposition creates more overhead than value.
AI Engineering
Different AI tasks can justify different models when quality, cost and data constraints vary. Routing should follow task-specific tests, not an assumption that one model always wins.
AI Architecture
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.
Keep in touch
New essays arrive as they publish. Every so often there is a shorter note when something in the work is worth passing along. If you would rather not do email, the same posts go out on the RSS feed.
Done. The next essay will show up when it publishes.
No essays in this topic yet.