Skip to content
WiTHUSiN

[ TECHNOLOGY ]

Four technologies that put LLMs to work in real business

WiTHUSiN builds technology that verifies what LLMs produce, connects multiple models, and automates work end to end. We run our own products on it and apply it to our clients’ operations.

01 / Verification

AI Output Verification

RequestPrompt · PR descriptionReq. 1Apply couponReq. 2Reject unknownReq. 3Save to orderCheckRequest ↔ codeAs requiredDiffersNo code
Request → Requirements → Check → Verdict

The problem

The faster AI writes code, the more checking that code against the request becomes the bottleneck in development.

Our approach

We turn the request given to the AI and the PR description into a list of requirements, then check each one against the changed code. Builds, tests, security and dependency checks run automatically and repeatedly, and every error is classified by type.

Components

  • Intent extraction that pulls requirements out of the request
  • Verdicts that match each requirement against the code
  • Multi-stage automated checks combining builds, tests, security rules (Semgrep) and dependency existence checks

In use

  • DevHive

    A GitHub PR tool that checks AI-written code against your request.

In use: DevHive

Check that AI-written code does what you asked.

The faster AI writes code, the more checking that code against the request becomes the bottleneck in development. DevHive turns the request given to the AI and the PR description into a list of requirements, and checks each one against the changed code. Anything the request left out becomes a question for the author, posted as a GitHub PR review.

PR #57 · Request ↔ AI output

Requirement 1

Add coupons to cart checkout.

As required
1function applyCoupon(total, code) {
2  const coupon = coupons[code]

Requirement 2

Don’t discount unregistered ones.

Differs
3  return total - (coupon?.amount ?? 1000)

Unregistered coupons get 1,000 off

Not stated · Ask the author

What if the coupon is bigger than the order?

Send

PR #57 · Review

Sent to PR #57

devhiveWhat if the coupon is bigger than the order?

VerdictsAs requiredDiffersNo code
Example: Your request ‘Don’t discount unregistered coupons’ links to line 3 of the code, which gets a ‘Differs’ mark; then a question the request never answered, ‘What if the coupon is bigger than the order?’, is sent to the PR as a review

How it works

  1. 01

    Gather requests

    Pulls requirements out of prompts, the PR description and the author’s answers, one per row.

  2. 02

    Check request vs. code

    Each requirement gets a verdict: ‘As required’, ‘Differs’ or ‘No code’.

  3. 03

    Send questions

    Asks the author about cases the request didn’t cover, and posts the questions as a PR review.

What it keeps to

  • It doesn’t run or change your code.
  • Only what a person sends is posted to the PR.
  • People approve and merge on GitHub.

Status

Private beta. We’re opening to invited teams first.

Visit DevHive(opens in a new window)

02 / Orchestration

LLM Orchestration

QuestionUser requestRoutingModel choiceModel AModel BModel CAnswerContext kept
Question → Routing → Model A/B/C → Answer

The problem

Without good prompts you can’t get the most out of a model, and every time you switch models you have to start the conversation over.

Our approach

We connect models like ChatGPT, Claude and Gemini into a single workflow. Switch models without losing context, and refine prompts to cut token costs.

Components

  • Multi-model connection
  • Model switching that keeps context
  • Automatic prompt optimization
  • Token savings
  • Bring your own API key (BYOK)
  • Knowledge base

In use

  • AI Weave

    A workspace for getting the most out of ChatGPT, Claude and Gemini in one place, no prompt skills needed.

03 / Agents

Agent Automation

InterviewGather needsDocumentsPRD · specsBuildAuto buildDeployAuto deploy
Interview → Documents → Build → Deploy

The problem

Requirements are scattered across people’s heads and many places, and getting from documents to a working result means carrying them over by hand at every step.

Our approach

AI asks questions to draw out requirements, then builds step by step from plans and specs to a working result. It connects to coding agents through MCP.

Components

  • Interview-style requirement gathering
  • Automatic planning documents (PRD, feature specs, user flows, test contracts)
  • Build and deployment automation
  • MCP integration

In use

  • MUOT

    Talk with AI and get from a plan to a working web service.

  • Backier

    A tool for designing APIs and testing them on the spot.

  • DevHive

    Asks the author about anything the request left out, and posts it as a PR review.

04 / Multimodal

Multimodal Understanding

VoicePhotoContextWorkRecordsFarming logsDocumentsCertifications
Voice · photos + work context → Records · documents

The problem

Information in the field is left scattered in voice notes and photos, and someone has to turn it into work documents all over again.

Our approach

We understand voice and photos together with the context of the work. Information scattered across the field becomes records and documents ready to use.

Components

  • Records from voice and photos
  • Automatic work documents (farming logs, GAP and organic certification paperwork)

In use

  • famfam

    An AI farming log app: record by voice and photo, and your farming log and certification documents are written for you.

Have a problem you want to solve with AI? Let’s talk.

Talk to us