Skip to content
The codeclawd steel claw holding an amber task above its circular platform

Agentic engineering, full-stack development, training and AI strategy

Intelligence.Put to work.

We build AI agents, automate knowledge-heavy work, and ship full-stack products. Your knowledge. Your tools. Your team in control.

Start a projectExplore the work

Built, then run every month by the people who built it.

YOUR KNOWLEDGE
+COORDINATED ACTION

Each reply to a comment on a drawing took 30 to 45 minutes. Most of them said the same thing.

Ten years of projects sat behind those replies.

One scholar answered the same question for years.

Tax preparers chased their clients for documents by email, until the intake became one upload per client and one file per return.

Not another demo. A working system.

Real retrieval, running in your browser. The records are invented; the search is not. Ask it something the archive cannot answer and it will say so.

Drop in a customer's comment on a drawing. The agent finds the reply you already gave and drafts the next one. Two comments to try. The records answer one of them.

42 records · 126 chunks

Made-up records, real retrieval. Nothing sends until the engineer approves it.

Draft ready for the engineer to approve

Customer asks whether the flange laminate on the tank drawing can be reduced

Thanks for the comments. The flange laminate stays as drawn, because the bolt load governs it and not the wall. We can take one ply out of the tank wall if you accept the revised deflection, and we will reissue the drawing at the next revision.

Reply 2019-07 · project 12 · drawing 2041 rev C

The flange laminate holds at the thickness shown on the drawing because the bolt load governs it, not the wall.

How it holds up.

From the records to a reply a person approved: the stages every answer passes through, drawn as we build them.

Ingestrecords

Chunkpages

Embedexcerpts

Retrieveboth ways

Rerankshortlist

Citeor refuse

Humangate

Only thensend

records to vectorstwo lists, one ordercited, approved, sent

Retrieval and approval pipelineOrthographic. Not to scale.

Ingestrecords

Chunkpages

Embedexcerpts

Retrieveboth ways

Rerankshortlist

Citeor refuse

Humangate

Only thensend

records to vectorstwo lists, one ordercited, approved, sent

General noteChecked against gold answers written from the source documents, never from the output of the system itself.

Ingest the records
Letters, drawings, scanned files and an inbox come in as they are. The originals stay read only.
Chunk the pages
Each page is cut into excerpts small enough to quote and to cite by page.
Embed with one multilingual model
One multilingual model turns each excerpt into a vector, so a question in one language finds a record written in another.
Retrieve, dense and lexical together
Two searches run on every question, one by meaning and one by exact words, and their results are merged into one list.
Rerank the candidates
A second model reads the merged shortlist and reorders it, so an off topic excerpt that scored well drops below the one that answers.
Cite the page, or refuse
Every answer names the page it came from. When the records do not cover the question, the system says so instead of answering.
Human approval gate
A draft waits for a person. Nothing is exported or sent until the engineer or the owner signs it off.
Only then, send
The reply goes out only after that approval, from the drafts folder, in the voice of the engineer who approved it.

What we build with.

Independent use only. No affiliation, partnership, sponsorship or endorsement implied.

What we do.

Four kinds of work. Most engagements draw on more than one.

  1. Agentic engineering

    Agents that read your records, draft the reply or the document, and stop at a person before anything sends. Retrieval with citations, refusal when the records do not answer, and an approval gate on every action that leaves the building.

  2. Full-stack development

    Web apps, phone apps, and internal tools, built around the way you already work. The data model, the interface, the hosting, and the accounts, all in your name from the first commit.

  3. Training and enablement

    Working sessions for the people who will run the system and for the people deciding what to build next. We teach the mechanism, not the hype, so your team can judge a demo, read a log, and say no to a vendor.

  4. AI consulting and strategy

    Which task to automate first, which model to trust with it, and where the data may live. A written plan with a fixed price at the end, and sometimes the honest answer that you do not need us.

Three ways to work with us.

Bring the process that takes too long, the product you want to ship, or a plan that needs a second opinion.

  1. Audit and plan

    Work out what to build first.

    You bring the process that eats the most time. We map it, test where software or a model helps, and hand you a build plan with a fixed price in writing. Sometimes the answer is that you do not need us.

    One to two weeks. A plan and a price.

    Bring a process
  2. Custom software and AI systems

    Build the system.

    Knowledge systems that answer from your records, internal tools, and web and phone apps for your customers. We build around the way you already work. The code and the accounts are in your name from day one.

    Four to eight weeks for the first workflow.

    Describe a build
  3. Monthly operation

    Keep it running.

    New records go in. We correct wrong answers, replace retired models, and build the next useful feature. The people who built it are the people on the call.

    Monthly. A report you can read in five minutes.

    Ask about the retainer

What we have built,without the client names.

COMMENTRECALLDRAFT

Operated

Agentic automation

The drawing-comment agent

  • Ten years of prior replies, indexed by project and drawing
  • Drafts the reply in the engineer's own voice
  • Three approval gates, and it never sends on its own

A customer's comment on a drawing arrives, the agent finds the reply the shop already gave, and the engineer signs the draft or changes it.

QUESTIONRULINGCITE

Live

Question answering

The rulings assistant

  • Questions in two languages over one closed corpus of published rulings
  • Dense and sparse retrieval together, then a reranking pass
  • Says it found no ruling rather than inventing one

People had been asking one scholar the same questions for years. The assistant answers from his published rulings and names the ruling, or says the rulings do not answer it.

CHATEXTRACTFILE

Live

SaaS product

The tax intake SaaS

  • A conversation collects the intake instead of a form
  • Typed tools write each field into the database as it talks
  • One upload per client, one file per return

Tax preparers were chasing their clients for documents by email. Clients now answer once, the app checks what is missing, and the preparer opens a complete file.

SEARCHPASSAGECITE

Live

Research platform

The classical text platform

  • Semantic and lexical search in one database, not a stack of stores
  • A trilingual corpus with the chain of transmission beside the text
  • Every answer names the passage it was taken from

A library of classical texts, searchable by passage, where every result names the edition and the page.

ARCHIVEINDEXANSWER

Operated

Knowledge system

The records question box

  • Every page indexed by job, date and document number
  • Drawings and calculations read as pages, not as typed text
  • Refusal by default when the records do not answer

A fabrication shop kept its pricing history on paper. We built a question box that answers with the page it came from.

Read the case study

BRIEFREHEARSEHIRED

Delivered

Preparation

AI interview preparation

  • One candidate, prepared for an interview on AI systems
  • Rehearsed against the technical ground the role stands on
  • The candidate was hired as a chief information officer

We prepared one candidate for an AI interview, and the candidate was hired as a chief information officer.

Position 1 of 6

What we have done.

Three systems, described without client names. The counts are theirs.

  1. Knowledge systemOperated

    The records question box.

    A fabrication shop kept its pricing history on paper. We read the archive page by page, drawings and calculations included, and built a question box that answers with the page it came from. Lookup went from an afternoon to minutes.

    671
    letters read and indexed
    1,121
    pages, drawings and calculations included
  2. Question answeringLive

    The rulings assistant.

    People had been asking one scholar the same questions for years. The assistant answers from his published rulings in either of two languages, names the ruling it answered from, or says that no ruling answers the question.

    Two
    languages over one closed corpus
    One
    reranking pass after dense and sparse retrieval
  3. SaaS productLive

    The tax intake SaaS.

    Tax preparers were chasing their clients for documents by email. Clients now answer once in a conversation, typed tools write each field into the database as they talk, and the preparer opens a complete file.

    One
    upload per client
    One
    file per return

Questions owners ask us first.

How an engagement runs, what you get, and what stays in your name.

Both. Businesses hire us for internal systems and customer-facing apps. Founders hire us for products. Scope follows the people who will use it.

It follows the size of the job. The audit ends with a fixed build price in writing, and you can stop after any stage. We have no package prices to squeeze your project into.

The audit takes one to two weeks. A first build usually ships four to eight weeks after that. Messy records take longer, and we tell you so before you commit.

Still have a question?

Bring it to the first call. We answer most of them in the first half hour.

Start a project

What takes too long in your business?

Tell us the process that eats the day, the product you want to build, or where you are stuck. One concrete problem is enough to start.

This form builds a brief you can copy and keep. It sends nothing and stores nothing.

Built, then run every month by the people who built it.

Start a project