01
Alive
AI that gradually learns how you work.
It draws reusable habits out of your past work — the ones you have confirmed — so every new task starts from experience you already have.
From generating content to getting work done
AI that understands your work,
and moves it forward.
Huaxin Technology is building three AI work systems: one that learns the working habits you have confirmed, one that makes every tender requirement traceable, and one that keeps a team's knowledge and decisions out of individual memory.
We believe good AI does more than answer questions — it becomes part of how work actually gets done.
02
Inkstone
Every tender requirement, answered.
It parses the tender clause by clause, checks the responses against each one, and flags what still needs confirmation.
03
Twin
A team that remembers its own decisions.
It connects scattered project files and past agreements, so every answer can be traced back to the original record.
01 — Personal work intelligence
Alive
AI that does not just recognise the task, but gradually understands how you work.
You should not have to explain your preferences from scratch every time.
Alive looks for reusable experience in your past work; once you confirm it, that experience carries into the next task.
Why Alive?
General AI can produce fluent text without knowing what you actually need.
It may get the company name wrong, ignore the data layout you are used to, or bring back phrasings you have already rejected.
The question is not whether AI writes well, but whether it can work to your standards.
How Alive works
- 01 Find relevant experience Search past records and confirmed habits for material related to the current task.
- 02 Extract working methods Identify reusable practices in real work output — not just statistical features of style.
- 03 Assemble the capabilities needed Work out which tools, skills and processing steps this task requires.
- 04 Produce a draft for review Generate a draft from the relevant experience, and state plainly what a person needs to check.
- 05 You decide what is worth remembering Candidate habits do not become long-term guidance automatically. They enter future work only after you confirm them.
Experience becomes the starting point for the next task
The value of a piece of work should not end when the document is finished.
Alive aims to settle verified experience into the system, so repeated work shrinks and the AI gets closer to how you actually work.
What it offers
- Traceable experience: important working habits come with a source of evidence.
- Memory you confirm: the model's guesses are never treated as your preferences.
- A visible process: from retrieving experience to producing a draft, the steps stay explicit.
- Results you sign off: the AI prepares, a person checks and decides.
02 — Tender document processing
Inkstone
Complex tender requirements, turned into responses that can be checked one by one.
The hard part of a tender is not writing. It is understanding every requirement accurately, keeping the response complete, and finding gaps in evidence and open questions in time.
Inkstone builds a structured response process around the tender document and the company's existing material, so every requirement has a status and every draft keeps its basis for review.
In tendering, the expensive mistake is omission
A document can read well and still carry risk — a missed clause, a missing proof, a requirement understood the wrong way.
When a team has to page through long documents, check clauses by hand and hunt for past material, time and omission risk grow together.
How Inkstone works
- 01 Parse the tender document Extract the structure, sections and specific requirements into a clear clause list.
- 02 Build a requirement-by-requirement response index Link each requirement to its response, its supporting material and its processing status.
- 03 Identify gaps and uncertainties Flag what is unanswered, under-evidenced or still to be confirmed — never treating an unknown as done.
- 04 Draft the responses Organise content from the tender requirements and existing company material, keep it marked as a draft, and hand it to a professional to review.
- 05 Support human checking Keep clauses, responses and source material aligned so a reviewer can find problems faster.
Prove reliability before chasing generation
Inkstone's research focus is whether the system can parse documents accurately, spot response gaps, and hold back content that lacks a basis.
Capability advances step by step. Only a feature that has been tested against a preset standard moves to the next stage.
Design principles
- Check item by item: a document that looks complete never replaces checking the individual clauses.
- Traceable basis: link responses to the original requirements and supporting material wherever possible.
- Visible uncertainty: missing material and open questions are marked explicitly.
- A person keeps the final call: generated content stays a draft until the owner reviews it before submission.
03 — Enterprise knowledge collaboration
Twin
A team's shared memory should not depend on one person's memory.
Project material, delivery agreements, past decisions and work records usually sit in different files and different hands. Over time the same question is confirmed again and again, and the same decision is understood in different versions.
Twin connects that scattered information so a team can work from records it can trace.
Turning repeated questions into something on record
How did we decide this last time?
What is the latest delivery requirement?
Which document did this conclusion come from?
When the answers sit in different versions of the material, the team has to keep searching, checking and confirming.
Twin is not about piling files into one place; it is about keeping origin, version and relationships intact as the information changes.
How Twin works
- 01 Connect the material you choose A baseline is taken from the locations you specify — nothing is collected by default.
- 02 Detect what changed Changes against the existing baseline are recorded, so the team can see what has been updated.
- 03 Keep origin and version Syncing preserves the source marker and version information that later queries depend on.
- 04 Answer from the original records A query returns the answer together with the passage it came from, so the reader can check it.
- 05 Keep up as material changes When the material changes, later queries build on the updated information.
From syncing files to shared understanding
Twin's aim is a team that does not have to hold collaboration together by word of mouth.
Information can be updated, decisions can be traced, answers can be checked. Even as members change, important agreements need not rest on one person's memory.
What it offers
- Incremental updates: watch what changed rather than reprocessing everything.
- Traceable origin: answers link back to the corresponding original records.
- Versioned evidence: users can tell which version a piece of information came from.
- Permission-scoped: access follows the actual design and configuration.
- Reliable sync: recovery after a dropped connection and duplicate records are part of what gets verified.
Principles
How we make AI worth trusting
Three systems, four shared principles.
AI can take on more work, but capability should not be bought with verifiability or with a person's right to decide.
We are not building an unconstrained automation system, but a way of working that can be verified step by step, improved continuously, and leaves the key decisions in human hands.
-
01
Draft first
Offer something reviewable before deciding whether to use it.
Generated content starts as a draft. Only after reading, checking and the necessary edits does an owner decide whether to use it for real.
-
02
Judgement stays human
Let AI support decisions, not carry the responsibility.
Where professional judgement is needed, where commitments are involved, or where consequences cannot be undone, the system keeps an explicit human review and sign-off step.
-
03
Everything citable
Every important conclusion should be traceable to a basis wherever possible.
When an answer comes from a document, a record or a past agreement, the system provides that source. Without sufficient basis it states the uncertainty plainly, rather than covering the gap with fluency.
-
04
Autonomy is earned
How much is automated is decided by verification results.
One success does not grant more authority. Wider automation is considered only after a task type keeps meeting its preset standard and has been verified by people often enough.
How we judge our own work
We care not only whether AI can finish a task, but whether it can show its basis, surface its problems, and stop when it is not sure.
To us, the intelligence that matters is not being capable of everything — it is knowing what you can do, on what basis, and when to hand the decision back.
Research notes
Research notes
We record not only what progressed, but also the assumptions testing killed.
AI reliability cannot be proven by a demo. We record test results, the problems they exposed and the design changes they caused, so capability rests on actual evidence.
Similar style is not authorship
In a test set containing 13 AI-written pieces, every single one passed the existing style baseline.
That showed the method could not separate similarity of style from authorship. We moved style checking to a different place in the pipeline as a result.
Too few samples means no conclusion yet
We tried to classify edit intent automatically from the differences between drafts and human revisions.
The samples did not support a reliable judgement, so the capability stays marked as unverified rather than passed.
Do not carry English assumptions into Chinese
While processing Chinese text, we found that a whitespace-tokenisation assumption had silently mis-handled about 65% of the corpus.
Only after fixing that foundation was it worth evaluating the model and the pipeline on top of it.
How we treat failure
Every test should be able to change our mind.
If the evidence does not support an assumption, we change the plan. If the samples are too few, we mark it unknown. If a feature has not met its standard, we do not describe it as finished.
The point of keeping a research record is not to prove we never get it wrong, but to make sure an error can be found, explained and corrected.
Contact
Start from a real problem at work
We are exploring how AI can take part more reliably in personal work, tender processing and enterprise knowledge collaboration.
If your work involves repeatedly sorting material, checking requirements one by one, hunting for past decisions or maintaining your own working standards, we would like to hear about the real situations.
Let us start from a concrete problem, not a grand AI concept.