You will get a knowledge base every employee's own AI can read, with permissions
Project details
Everyone at your company already uses AI. Each of them has a private chat history that walks out of the door with them, and none of it can answer a question about your own contracts, prices or procedures.
This fixes that without anyone sharing an account or changing tools.
People read it in a browser. Their own AI reads it as native tools: MCP for Claude and Cursor, an OpenAPI action for ChatGPT. Same check either way.
PERMISSIONS THAT HOLD
Two gates, both server-side, so a leaked token cannot read past its owner: folder read scope, plus a per-document role floor so one file in a shared folder stays restricted. The folder a document sits in IS its permission scope, so there is no second place to configure and forget.
REAL DOCUMENTS
Imports .docx, .xlsx, .pptx, .pdf, .md and .csv. Diacritic-folded search, so typing without accents still finds accented text. Answers return passages and source ids, never generated text, and when nothing matches it says so instead of inventing an answer.
WHAT IT COSTS TO KEEP
Nothing per person, ever. Node 18 and nothing else: no database, no vector store, no npm install. Documents stay as markdown on disk, so the knowledge outlives the software.
This fixes that without anyone sharing an account or changing tools.
People read it in a browser. Their own AI reads it as native tools: MCP for Claude and Cursor, an OpenAPI action for ChatGPT. Same check either way.
PERMISSIONS THAT HOLD
Two gates, both server-side, so a leaked token cannot read past its owner: folder read scope, plus a per-document role floor so one file in a shared folder stays restricted. The folder a document sits in IS its permission scope, so there is no second place to configure and forget.
REAL DOCUMENTS
Imports .docx, .xlsx, .pptx, .pdf, .md and .csv. Diacritic-folded search, so typing without accents still finds accented text. Answers return passages and source ids, never generated text, and when nothing matches it says so instead of inventing an answer.
WHAT IT COSTS TO KEEP
Nothing per person, ever. Node 18 and nothing else: no database, no vector store, no npm install. Documents stay as markdown on disk, so the knowledge outlives the software.
AI Algorithms
Large Language ModelAI Applications
Natural Language UnderstandingAI Models
ChatGPT, GPT-4What's included
| Service Tiers |
Starter
$180
|
Standard
$590
|
Advanced
$1,200
|
|---|---|---|---|
| Delivery Time | 1 day | 5 days | 12 days |
Number of Revisions | 0 | 2 | 3 |
AI Model Integration | - | ||
Batch Normalization | - | - | - |
Database Integration | - | - | - |
Detailed Code Comments | |||
Image Upscaling | - | - | - |
MLOps | - | - | - |
Model Deployment | - | - | - |
Model Documentation | - | - | - |
Model Monitoring | - | - | - |
Model Testing & Optimization | - | - | - |
Model Tuning | - | - | - |
Natural Language Processing | - | - | - |
NLP Tokenization | - | - | - |
Pre-Training | - | - | - |
Prompt Engineering | - | - | - |
Setup File | |||
Source Code |
Frequently asked questions
About TJ
AI Infrastructure - Internal Knowledge Bases & LLM Cost Control
Hanoi, Vietnam - 2:53 am local time
I build the layer that fixes that, and I ship it as software you own and run yourself rather than a subscription you rent.
WHAT I BUILD
Internal knowledge bases with real permissions. Staff read them in a browser; their own AI reads them as native tools over MCP (Claude, Cursor) or an OpenAPI action (ChatGPT). Two permission gates are evaluated server-side, so a leaked token still cannot read past its owner. Handles .docx, .xlsx, .pptx and PDF, and searches accented languages correctly.
LLM cost control gateways. One vendor contract, one API key per person, model access tiered by role, and a rolling budget enforced before the request goes out rather than discovered on the invoice. Plus the admin dashboard finance needs to reconcile against the vendor's own console.
The automation around both: marketplace and ERP integrations, scheduled reporting, document pipelines, and browser automation for platforms with no usable API.
HOW I WORK
Fully asynchronous. Written updates, no standing meetings required. Every change ships with a backup and a before/after comparison, and I report what broke as plainly as what worked.
I will not quote you a savings percentage before seeing your usage, and I do not resell AI access. Everything routes through your own vendor account under your own contract.
Stack: Node.js, Python, SQLite, Linux, Nginx, Cloudflare, Firebase, Vercel.
I would rather show you a working system than describe one. There are live demos in my portfolio. Click into them, they run in your browser.
Steps for completing your project
After purchasing the project, send requirements so TJ can start the project.
Delivery time starts when TJ receives requirements from you.
TJ works on your project following the steps below.
Revisions may occur after the delivery date.
Permission map, then deploy
We turn your departments into folders and role floors first, because the folder a document sits in IS its permission. Then it goes up on your server, over HTTPS, with accounts created.
Import, connect the AI tools, then hand over
Your documents go in and get indexed. Claude and Cursor connect over MCP, ChatGPT as an OpenAPI action. Then I test the same query from two different roles in front of you, so you can see the permission gates actually hold.