AI virtual machine (on local Server without GPU). Target: Financial Data Analysis
Budget: €30 – €250 EUR
NO PLACEHOLDERS. Do not insert a placeholder: please read the project details carefully because here are all the details you need to make an accurate estimation.
DON’T USE AI to read this announcement: AI switches off your intelligence and makes you miss out on many details that you can only pick up by reading my request carefully.
We are looking for an experienced AI consultant to build a working Proof of Concept (PoC).
PoC means few time and low budget: if PoC success, linked project will start.
The goal is to create a locally deployable AI system (on-premise, on Windows o.s.), PRE-TRAINED in the financial domain, for example
* FinGPT of Qwen --> https://github.com/ai4finance-foundation/fingpt;
* FinBERT on Hugging Face --> https://huggingface.co/ProsusAI/finbert;
* FinText --> https://huggingface.co/FinText;
* BloombergGPT
* ChatGPT for Finance;
* DeepSeek-R1;
* FinRobot --> https://github.com/ai4finance-foundation/finrobot;
* ...
IF the model requires fine-tuning, it is important to agree on a training plan.
Please pay attention on environment performances.
The aim is to create
- a system capable of reading large PDF files (>300 pages) containing tables (see attachment). PDF files contain text only. The tables differ in terms of the data they contain, their structure, the number of fields, the number of rows and the subject matter covered;
- an LLM system that can be deployed locally (on-premise) on Windows servers (no WSL - Windows Subsystem for Linux) WITHOUT a GPU, PRE-TRAINED in the financial area and capable of analysing the data read from the aforementioned PDF files;
- a method of access both via web (such as a Open WebUI) but primarily via APIs (to be written in .py or the language of your choice)
Full opensource.
The Windows VM (VMware or VirtualBox) must be ready to use. HW available: 16gb RAM and 250gb HD and 2 cores.
Accuracy and the absence of hallucinations are very important
Ensure the system operates fully offline, independent of any external cloud services.
Please note that considerations for High Availability (HA), Disaster Recovery (DR), or Load Balancing are not required. It's only a PoC.
I will accept the assigned job when I am able to:
• Start the VM and upload a PDF to generate a detailed answer;
• Recreate the environment from scratch by following the instructions provided.
----- How will it be used? -----
Mainly via API/WebService: applications (VB/ASP/ASP.NET) will collect requests from the connected user and request answers from local AI service via APIs.
However, a web interface (such as ChatGPT, Claude, ...) is required to make impromptu requests
----- Tests already completed -----
Regarding data read from PDFs and analysing it, much of preliminary tests were carried out using free online models/services, and the results were as follows:
- GPT-4o fails
- GPT-5-mini fails
- Mistral fails
- AI Studio (Google) OK
- Deepseek-chat-3.2 fails
- scispace.com OK
- super.myninja.com OK
----- Budget -----
Please, no ask me «what is your budget for this project?»
Bonuses provided at the end of the project for compliance with the timing and quality of the results.
No upfront. No payment before successful completion of all final test.
The proposal must cover the entire project, not just a "minimum viable product" (VMP).
----- Milestones -----
The project is a black box: either everything works or nothing works, so it isn’t possible to break it down into milestones
----- Collaboration -----
The consulting engagement will commence in June and will be conducted full-remote.
Participation in daily update meetings is mandatory and non-negotiable. Failure to adhere to this requirement will result in immediate project disengagement, without exception.
----- FAQ -----
1. Which exact open-source foundation model(s) should I standardize on for the offline LLM (e.g., FinGPT/Qwen, Mistral-based GGUF)? Do you need table-aware behavior, or is pdfplumber-based extraction sufficient?
--> From my tests, extraction with pdfplumber fails. It's not 100% reliable.
2. For the API outputs, do you want structured JSON (table fields, confidence, citations by page/table id) alongside text, or only text?
--> I need 100%. If the PoC doesn't provide guarantees, the project won't start. The confidence level must be maximum. I expect the JSON to contain the table title, column titles, and the text contained in the tables. Check the format diversity in the attached example.
3. Since the PDFs exceed 300 pages and contain highly variable financial tables, what level of accuracy do you expect for table extraction and reasoning—should the system only answer questions from extracted content, or also perform financial analysis such as trend detection, ratio calculations, risk assessment, and cross-document comparisons?
--> step 1 = no financial analysis. Only text extraction 100% accuracy.
--> step 2 = financial model (as suggested) could answer to user prompts.
4. Do your PDFs follow any semi-consistent format (same issuer/report type), or are they completely heterogeneous?
--> No, data format are similar but not identical.
5. What does a passing answer look like for your final test: narrative summary, extracted figures with traceability, cross-table comparisons, or exact answers tied to page and table references?
--> Accurate and perfect extraction of data into tables.
6. For the final validation, will you compare the JSON output against a known reference dataset, or will the verification be performed manually against the original PDF tables?
--> Test performed manually: it's the only way to verify 100% accuracy.
DON’T USE AI to read this announcement: AI switches off your intelligence and makes you miss out on many details that you can only pick up by reading my request carefully.
We are looking for an experienced AI consultant to build a working Proof of Concept (PoC).
PoC means few time and low budget: if PoC success, linked project will start.
The goal is to create a locally deployable AI system (on-premise, on Windows o.s.), PRE-TRAINED in the financial domain, for example
* FinGPT of Qwen --> https://github.com/ai4finance-foundation/fingpt;
* FinBERT on Hugging Face --> https://huggingface.co/ProsusAI/finbert;
* FinText --> https://huggingface.co/FinText;
* BloombergGPT
* ChatGPT for Finance;
* DeepSeek-R1;
* FinRobot --> https://github.com/ai4finance-foundation/finrobot;
* ...
IF the model requires fine-tuning, it is important to agree on a training plan.
Please pay attention on environment performances.
The aim is to create
- a system capable of reading large PDF files (>300 pages) containing tables (see attachment). PDF files contain text only. The tables differ in terms of the data they contain, their structure, the number of fields, the number of rows and the subject matter covered;
- an LLM system that can be deployed locally (on-premise) on Windows servers (no WSL - Windows Subsystem for Linux) WITHOUT a GPU, PRE-TRAINED in the financial area and capable of analysing the data read from the aforementioned PDF files;
- a method of access both via web (such as a Open WebUI) but primarily via APIs (to be written in .py or the language of your choice)
Full opensource.
The Windows VM (VMware or VirtualBox) must be ready to use. HW available: 16gb RAM and 250gb HD and 2 cores.
Accuracy and the absence of hallucinations are very important
Ensure the system operates fully offline, independent of any external cloud services.
Please note that considerations for High Availability (HA), Disaster Recovery (DR), or Load Balancing are not required. It's only a PoC.
I will accept the assigned job when I am able to:
• Start the VM and upload a PDF to generate a detailed answer;
• Recreate the environment from scratch by following the instructions provided.
----- How will it be used? -----
Mainly via API/WebService: applications (VB/ASP/ASP.NET) will collect requests from the connected user and request answers from local AI service via APIs.
However, a web interface (such as ChatGPT, Claude, ...) is required to make impromptu requests
----- Tests already completed -----
Regarding data read from PDFs and analysing it, much of preliminary tests were carried out using free online models/services, and the results were as follows:
- GPT-4o fails
- GPT-5-mini fails
- Mistral fails
- AI Studio (Google) OK
- Deepseek-chat-3.2 fails
- scispace.com OK
- super.myninja.com OK
----- Budget -----
Please, no ask me «what is your budget for this project?»
Bonuses provided at the end of the project for compliance with the timing and quality of the results.
No upfront. No payment before successful completion of all final test.
The proposal must cover the entire project, not just a "minimum viable product" (VMP).
----- Milestones -----
The project is a black box: either everything works or nothing works, so it isn’t possible to break it down into milestones
----- Collaboration -----
The consulting engagement will commence in June and will be conducted full-remote.
Participation in daily update meetings is mandatory and non-negotiable. Failure to adhere to this requirement will result in immediate project disengagement, without exception.
----- FAQ -----
1. Which exact open-source foundation model(s) should I standardize on for the offline LLM (e.g., FinGPT/Qwen, Mistral-based GGUF)? Do you need table-aware behavior, or is pdfplumber-based extraction sufficient?
--> From my tests, extraction with pdfplumber fails. It's not 100% reliable.
2. For the API outputs, do you want structured JSON (table fields, confidence, citations by page/table id) alongside text, or only text?
--> I need 100%. If the PoC doesn't provide guarantees, the project won't start. The confidence level must be maximum. I expect the JSON to contain the table title, column titles, and the text contained in the tables. Check the format diversity in the attached example.
3. Since the PDFs exceed 300 pages and contain highly variable financial tables, what level of accuracy do you expect for table extraction and reasoning—should the system only answer questions from extracted content, or also perform financial analysis such as trend detection, ratio calculations, risk assessment, and cross-document comparisons?
--> step 1 = no financial analysis. Only text extraction 100% accuracy.
--> step 2 = financial model (as suggested) could answer to user prompts.
4. Do your PDFs follow any semi-consistent format (same issuer/report type), or are they completely heterogeneous?
--> No, data format are similar but not identical.
5. What does a passing answer look like for your final test: narrative summary, extracted figures with traceability, cross-table comparisons, or exact answers tied to page and table references?
--> Accurate and perfect extraction of data into tables.
6. For the final validation, will you compare the JSON output against a known reference dataset, or will the verification be performed manually against the original PDF tables?
--> Test performed manually: it's the only way to verify 100% accuracy.