Insights
Thinking Out Loud
Essays on AI workflow engineering, automation, and data science, and the deep-technical work that proves the method transfers across domains.
-
AI Engineering
You Bought the AI Licenses. Here's Why Your Team Still Uses It Like a Chatbot.
Copilot, ChatGPT Enterprise, and Claude seats got approved months ago. Most of your team still opens a chat window, pastes something in, and copies the answer back out by hand.
-
Data Science
From Spreadsheet to Workbench: Automating the Data Assembly Around Your Modeling Software
HydroCAD, SewerGEMS, and Civil 3D are not the bottleneck. The hours your engineers lose reformatting spreadsheets before the model even opens are.
-
Data Science
Every Number Cited or Flagged: Building AI Workflows an Engineer Can Audit
AI output stalls at interesting and never reaches shippable for one reason: you can't trace the numbers. Here's what a provenance-clean workflow looks like, and why it's the difference between a demo and a deliverable.
-
AI + Tools
Keeping Client Data and IP Out of Your AI Tools: A Practical Boundary
The AI blocker at most science and engineering firms isn't a bad tool. It's a missing boundary that counsel can actually sign off on.
-
AI Architecture
How to Pick the First AI Workflow to Ship (and Which to Leave Alone)
Most teams pick their first AI workflow by how impressive it sounds. The teams that actually ship pick by how easily they can catch it being wrong.
-
AI Engineering
A Passing Test Is Not the Finish Line: Verifying AI Output Before It Ships
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
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.
-
Innovation
AI Doesn't Have an Energy Problem. We Do.
The Kardashev scale says a civilization's reach is gated by how much energy it can put to use, not by cleverness. A data center pulling enormous power off the grid isn't proof AI broke something. It's proof the technology works, and proof we've underinvested in energy for decades.
-
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 Engineering
Execution is the trivial part
AI made execution trivial, which broke the portfolio as evidence of skill. What differentiates now is judgment, and most of it never gets recorded. That is the difference between an agentic engineer and a vibe coder.
-
Federal Funding
A Federal Proposal Workflow Is Infrastructure
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
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.
-
R&D Strategy
Why Breakthroughs Start With What's Possible, Not What Customers Ask For
Capability expansion precedes need articulation. The most consequential technologies in history weren't solutions to articulated problems.
-
AI Engineering
The Knowledge Base Nobody Reads (Including You)
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
AI Didn't Break the Test. The Test Was Already Broken
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
The Federal Proposal Section No One Edits Twice
Action items, clarifying questions, and IP novelty get drafted once and never reviewed. That is where a large share of proposal losses originate, and where the team's confusion about scope, dependencies, and IP cleanly leaks through.
-
AI Engineering
Check the Brief Before You Trust the Build
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
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.
-
Biotech
Single-Cell Foundation Models: Three Adoption Lessons
Most new spatial transcriptomics projects reach for a single-cell foundation model, then hit the same three walls (preprocessing, hardware, regulatory) and discover the model was never the moat.
-
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.
-
Data Science
Six Friction Points in a Typical Drainage Study
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.
-
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.
-
AI + Tools
How to Give AI Current Web Data Without Exposing Your Files
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 ways to pressure-test SBIR Specific Aims
Five criterion-linked checks for pressure-testing whether SBIR Specific Aims make the innovation, approach, evidence, risk, and team logic easy to evaluate.
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
Why AI Coding Tools Should Run in a Sandbox by Default
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
Cloud LLMs Have Failure Modes Benchmarks Don't See
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
Your AI Doesn't Need to Read Your Whole Knowledge Base to Use It
Knowledge map traversal (navigating structured graphs on-demand instead of loading entire databases) is the superior pattern regardless of context window size.
-
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
The Hidden Tax of Splitting AI Work Across Too Many Agents
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
How AI Systems Remember What They Already Know
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 + Tools
Your AI's 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 Engineering
Routing the Same Task Through Four AI Models: What Each One Got Wrong
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
When Parallel AI Work Silently Corrupts Itself
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
How Small Should an AI Task Be?
When decomposing tasks for AI agents, there's a granularity floor: a point below which further decomposition creates more overhead than value.
-
AI Engineering
Why Most AI Work Doesn't Need Your Most Expensive Model
Most teams pick one LLM backend and stick with it. A better question is which tasks need the strongest model and which are better handled by a smaller or cheaper one.
-
AI Architecture
Writing, Reviewing, and Deciding Are Three Different Jobs
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.
No essays in this topic yet.