Contact person: Business Client

Phone:Show

Email:Show

Location: Reading, United Kingdom

Budget: Recommended by industry experts

Time to start: As soon as possible

Project description:
"This project is for phase 1 described below, if successful we can continue with phases 2-5 and related projects. Speed of delivery is important. The main deliverable will be a containerised application, cloud implemented (owner controlled). We will discuss and agree the appropriate cloud services. Let’s discuss what phase 1 will require and how long you will need to complete it.

Phase 1: I have approximately 300 texts (law journal articles, legal reports and court judgments) in PDF format that need to be individually analysed using a model like GPT-5 to compile a JSON object for each text. The analysis requires up to 10 separate prompts for each text, but the prompts are fixed and direct the model to output clean JSON. The task is to automate the prompt sequence and extraction of JSON answers for each text (PDF file). Parameters for phase 1 are:
• Browser accessible dashboard for managing multiple separate projects
• Each project to be separately configurable: load file, set API key(s), input and configure 1-10 fixed prompts for the project, and select LLM model from available and compatible models.
• Cost estimator: estimate cost for the project for each available and compatible LLM model before initiating the run.
• Prompt configuration for 1-10 fixed prompts:
o Fixed text content.
o Search the internet (how to achieve this through the API to be discussed).
o Linked prompts: Most prompts are self-contained (with the file), but configuration will need to have an option to link prompts for mimicking multi-turn conversation (eg the second prompt in the link needs to include the prompt-response content for the first prompt, etc).
o Staging: The output for the first prompt includes a citation for the text; the complete set of citations for all texts will optionally be used as part of the content for prompts 2-10. We need the ability to set stages for the prompt so that eg stage 1 can be run with only the PDF files and prompt 1; then the content of prompts 2-7 finalised for stage 2; with the possibility to add a stage 3 to run additional prompts later.
• Live progress: totals, succeeded/failed, tokens used/costs incurred vs estimate.
• The job must be pausable and resumable
• If a step fails after retries/repair, halt that file’s thread, log the failure, and continue with others. Resume/retry/modify controls for failed prompts/files.
• Persist provenance in the JSON: file name, uploaded file_id(s), page count (if computed), model ID, request IDs, and token usage (input/output/total).
• Concurrently run multiple prompts for faster output.
• JSON repair attempt on schema mismatch.
• Downloads: text-JSON, project-JSON, and zip folder all text-JSON and project-JSON together.
• Well-documented source code and infra templates (if applicable).

Phase 2: Automatic validation. Many of the JSON values are supposed to be direct quotes from the text, and the page number on which they appear. We need an automatic validation that searches for the quote, text matching, and confirms that it is real (and not a hallucination) and confirms the page number it appears on matches the value in the JSON. Will need to be able to configure the values for automatic validation from the texts, additional JSON keys for automatic validation state. Will need a second stage manual validation for failures. Report outputs TBD.

Phase 3: Manual validation. Other JSON values are supposed to be abstractions or conclusions. A sample (selectable by text) will need manual validation. Select text-JSON’s for manual validation, two frame view (PDF file left, JSON values right), step through JSON keys with Agree/Disagree/Flag (flag for tertiary review), record input to JSON and move to next key. Additional JSON keys for manual validation state. Report outputs TBD.

Phase 4: Consolidated reports, graphs and analytics.

Phase 5: Involves automating Google search and retrieval for 1000+ search strings (formulaic generation), and using a similar process to Phases 1-4 above to structure, validate and report.

My expectation is that you will develop the application using your own resources (eg Google Cloud, AWS, OpenAI API key). Once you have a working version for release you will implement it on my services (eg my Google Cloud instance and using my OpenAI API key) including configurations and continue to provide ongoing support and updates.

Based on past experiences with freelancers please note the following which are non-negotiable: I will not release any payment before the deliverable is received and has passed testing. Do not bid for this project if you do not have or are unable to obtain your own subscriptions for development. I will generally expect you to factor your costs into your rates, if necessary (rare exceptions) we can agree a budget for specific costs in advance – for your security I will prefund a milestone for the reimbursement that will be released only on receipt of your receipts/invoices for the costs actually incurred and completion of the deliverable." (client-provided description)