How the work actually gets built.
Detailed walkthroughs of the systems I build, the problems they solve and the decisions behind them. Where a project is a representative build rather than a named client engagement, it is labeled as such. I would rather show you exactly how I work than hand you a logo wall.
Government proposal compliance automation.
A small firm that writes proposals in response to government solicitations. Every solicitation is a long, dense document full of requirements that must each be tracked and answered, and missing one can sink a bid. The team was building compliance matrices by hand, a slow and error-prone slog on every opportunity.
These documents are messy and inconsistent, and the cost of a missed or misread requirement is losing the contract. A tool that is usually right is not good enough. It has to be checkable.
Scope to one bounded, provable job first. Read a solicitation, extract every requirement and build the compliance matrix automatically, with the output structured so a human can verify it fast.
An automated pipeline that watches for a new solicitation, reads it, pulls each requirement out and writes every one as a row in a structured compliance matrix, ready for the team to review and respond against. Alongside it, a set of assistants for pulling relevant past-proposal content and drafting sections against the firm's own style and rules.
The system is graded against hand-built answer keys so its accuracy is measured, not assumed. The output is built for human review rather than blind trust, and the drafting assistants work from the firm's approved past content and compliance rules rather than inventing claims.
Compliance matrix work that took a person the better part of two days collapses to minutes of automated extraction plus a fast human check. The team spends its time answering requirements instead of hunting for them.
Grounded internal knowledge assistant.
A property management operation with knowledge spread across policies, procedures and a few long-tenured staff. New hires could not find answers on their own and kept pulling senior people off their work.
The source documents were a mess. Two conflicting versions of the same procedure. A policy silently overridden by a newer email that never made it back into the official document. Sensitive personal data mixed into files that a general assistant must never surface. And one critical procedure that existed only in a veteran's head and nowhere on paper. An assistant built on that corpus without cleaning it first would confidently give wrong, unsafe answers.
Treat document cleanup as the majority of the job, because it is. Reconcile the conflicts, retire the overridden policy, wall off the sensitive data and identify the gaps where the honest answer is “this is not written down, ask a person.”
A retrieval-based assistant grounded strictly in the cleaned, reconciled document set, scoped to one team's real questions, that answers from the source material and points to where the answer came from.
The assistant is built to refuse rather than guess. Asked something outside its cleaned knowledge, it says it does not know and routes the person to the right human. Sensitive personal data is excluded from what it can ever surface. Refusal behavior is the feature that makes it safe to deploy, and it is tested for explicitly.
New staff get reliable answers on their own, senior people stop being a help desk and the organization's knowledge stops living in one person's memory. Just as important, the assistant is trusted because it is honest about the edges of what it knows.
This slot is reserved for the next named or anonymized client engagement. Real case studies drop in here as they close.
