user@tatmantech.com:~$ whoami▌
TatmanTech — 15+ years of enterprise and mid-market experience in systems,
operations, storage, disaster recovery, risk management, and technical
architecture, now centered on practical AI and machine learning integration. AI
and ML often bring new business paradigms, but that doesn't have to be a
frightening or disruptive process. Engagements expand or contract as needed,
drawing on outside expertise to grow with your needs.
We don't sell one-size-fits-all “AI transformation.” We meet the
challenge in front of you — adapting solutions to fit the structures you
already have, and adapting structures to fit the solutions you need —
whether that's a triage step, a translation layer, or something built to order.
// technique
This is a demonstrated technique, not a client case study — no client data or results are represented here.
Triage an inbox like an incident queue
Most small businesses handle every inbound message the same way: read them in
order, respond when there's time. An incident queue doesn't work that way —
severity gets classified first, and routing follows from that. The same instinct
applies to a sales inbox.
A lightweight classification step — run over new messages as they arrive
— can sort into four buckets: urgent, sales, support, spam. Urgent gets a
notification. Sales gets queued for follow-up. Support gets routed to whoever
owns it. Spam gets filtered without needing a dedicated tool. No new platform,
no new inbox — one classification step wired into what's already there.
This is a demonstrated technique, not a client case study — no client data or results are represented here.
Customer-message triage without buying a platform
A message carries information before anyone reads it — what language
it's in, how urgent it is, what it's actually asking for. Missing that signal
is the default: a five-alarm complaint waits in the same queue as a routine
question until a person gets to it.
Customer-service AI platforms are usually sold as a full replacement for how you
already work — new inbox, new dashboard, new monthly bill. Most of the
actual value is two smaller pieces: understanding what a message says regardless
of the language it arrived in, and flagging which messages need attention now.
Rather than replace what you already run, we build the connectors that let
differentiated tools work together — getting most of that value without
the security, risk, and process conflicts a full suite would introduce.
Unlike the writeups above, the links below point to real, public work — not a hypothetical example.
Dataset curation and structured prediction
Most “we don't have good data” problems are actually “we have
data, but no one has sorted it.” Curating and pruning records —
deduplicating, reformatting, discarding what's actually noise — is the
same groundwork a BI dashboard depends on to make a number mean what it claims
to mean. It doesn't need a BI platform; it needs someone to go through the data
and records carefully. That's dataset curation and construction as its own
discipline — and the exact same toolkit applied to your own records,
without needing a dedicated BI platform.
That work is public: python-code-dataset-500k [external],
a curated 500k-example dataset with 900+ downloads, and a five-dataset cluster
built from divergent sources — text, blog posts, tweets, and movie
reviews — around the same classification target:
myers_briggs_text_classify [external] and its siblings.
The same discipline extends to prediction:
a public example [external]
runs regression and classification models against structured data, then
generates synthetic data and validates it against the original — the
same approach behind demand forecasting, churn prediction, or anomaly flags on
your own records.
And to search: a public example [external]
runs a local similarity-search pipeline — dense and ColBERT embeddings, a
vector database, named-entity extraction cross-referenced across sources —
over a 600+ chapter, multi-tradition text corpus, browsable through a web app,
an API, and a CLI. The infrastructure behind it is real but extensive; what
matters here is that it's working, and built with known, current tools.
Like the writeup above, this links to real, public work.
Extending an existing system, not just building new ones
Most valuable systems already exist — the job is often extending what's
there, safely, rather than replacing it. That instinct comes from the same
place as the operations and architecture background above.
A public example: forking an existing open-source OSINT personal-intelligence
dashboard and using AI-assisted development to extend it well past the
original — the data-source count grew from 27 to 37+, alongside work on
new capabilities like an MCP server integration and cross-stream risk
correlation. See github.com/jtatman/Crucix [external].
Like the writeups above, these link to real, public work.
Multiple models arguing toward an answer
A single model gives you its first answer. Multiple models, prompted
differently and made to critique each other, tend to catch the same failure
modes a second reviewer catches on any other risk assessment.
socket-ai [external]
runs an arena where differently-weighted, differently-prompted LLM bots hold a
shared conversation, organized into topic-specific teams (planning, DevOps,
business development, and others) with full logging.
HegelianAI [external],
still under active development, pushes that further: multiple local models
argue a proposal through Hegel's thesis, antithesis, and synthesis, with a
provider-independent judge step and full tracing on every turn — a basic
looping structure, ahead of the self-improving kind.