Choose AI models
with evidence.

• A comparison of your current model and two lower-cost candidates on 50 supplied examples for one text task: classification, structured extraction or short summarization.
• Scoring rules agreed before testing, using your expected answers or acceptance rubric. Automated and model-judged scores are identified separately.
• Pass rates, observed API response times, token costs and cost per passing result under the stated checks.
• A manual review of 10 agreed examples across all three models, highlighting important regressions.
• A short recommendation: switch, run a larger test, or keep the current model.
• Raw outputs, results CSV and a Python script with the tested prompts and configuration. Where a candidate passes, its API example is included.
• Up to $15 of test API usage included. One clarification round within seven calendar days of delivery.

Scope: supported hosted text APIs; up to 2,000 input tokens and 500 output tokens per example, with reasoning budgets constrained where supported. No self-hosted GPU performance claims. Contact me before ordering if your models or task may exceed these limits.

Supported hosted APIs: OpenAI, Anthropic, xAI (Grok), and selected providers serving open models. Model availability and compatibility with your task are confirmed before testing.

• Your current model/provider, prompt and task description.
• 50 representative, non-sensitive examples in CSV or JSON, with expected outputs or a clear acceptance rubric. Identify your 10 most important examples.
• Required output format and the errors you cannot accept.
• Current API rates and representative monthly usage, if you want projected savings.
• Two candidate models, or permission for me to select supported alternatives within the test allowance.
• Confirmation that the examples may be processed by the selected hosted providers. Do not send passwords or API keys through the order form.

You can paste brief inputs into the request or provide a readable file link. Testing starts once the examples, options and scoring rules are complete.

Production integration, self-hosting, GPU rental, fine-tuning, autonomous agents, tool use, image/audio tasks, long-context testing, ongoing monitoring or unlimited prompt optimization. No guaranteed savings or equivalent quality outside the tested examples. Customer-specific private endpoints and proprietary model installations are excluded. Creating or labeling your evaluation dataset is not included.

• A comparison of two supported API-based extraction solutions, using open models where suitable.
• Testing on 20 supplied pages from one recurring document type, in one language, extracting up to 10 agreed fields per page.
• Field-level checks against your supplied answers, with missing, incorrect and invalid outputs shown separately.
• Observed processing times, test costs and estimated API cost per 1,000 similar pages under stated assumptions.
• A short recommendation, including document conditions that caused failures. If neither option is suitable, that is reported clearly.
• Extracted results, comparison CSV and a reusable Python API script with the tested extraction schema and setup instructions.
• Up to $15 of test API usage included. One clarification round within seven calendar days of delivery.

Scope: legible printed PDF, PNG or JPEG pages; one document family with similar layouts; 20 pages total, up to 20 MB. This is a focused pilot, not a production ingestion system. Contact me before ordering for unusual formats.

Supported hosted APIs: OpenAI, Anthropic, xAI (Grok), and selected providers serving open models. Model availability and compatibility with your task are confirmed before testing.

• 20 representative pages from one document type, with permission to process them through the selected hosted providers. Remove confidential or personal information before sharing.
• Up to 10 fields you need, their required formats and a reference answer for each field on every page. Mark absent fields explicitly.
• The mistakes that matter most: wrong totals, missing dates, incorrect identifiers or similar.
• Expected monthly page volume if you want a usage-cost projection.
• Two candidate solutions, or permission for me to choose supported options within the test allowance.

Provide the pages and reference answers through a readable file link. No API keys are needed in the order form. Testing starts once the files, schema and reference answers are complete.

Handwriting, document translation, bespoke model installation, training, GPU provisioning, creating reference answers, manual document cleanup, line-item tables with an unlimited number of rows, mixed document categories, production deployment or accounting/ERP integration. No guarantee of error-free extraction. Hosted API results do not establish self-hosted performance or privacy suitability.

• A comparison of three supported embedding models, including open models where suitable, using one shared retrieval setup.
• One corpus of up to 100 plain-text documents, 250,000 characters in total, tested against 30 queries with your relevance labels.
• A simple keyword-search baseline, so you can see whether embeddings improve these searches.
• The share of queries with a relevant document in the first five results, plus query-by-query results and missed documents.
• Observed indexing/query times and embedding usage costs in the tested environment.
• A short recommendation, raw rankings, CSV results and Python scripts to rebuild the index and rerun the comparison.
• Up to $10 of test API usage included. One clarification round within seven calendar days of delivery.

Scope: one language, one fixed text-splitting method and one retrieval configuration. Results measure retrieval, not the quality of a chatbot's generated answers. Contact me before ordering if your corpus needs extraction or custom infrastructure.

• Up to 100 plain-text documents in TXT, CSV or JSON, totaling no more than 250,000 characters, with stable document IDs.
• 30 realistic search queries and the IDs of the documents that should be relevant for each query.
• Your intended use and examples of searches that currently fail.
• Three candidate embedding models, or permission for me to choose supported options within the test allowance.
• Confirmation that the supplied content may be processed by the selected hosted providers. Remove confidential or personal information before sharing.

Provide a readable file link, or paste a small dataset into your request. Do not send passwords or API keys. Testing starts once the corpus, queries and relevance labels are complete.

PDF/OCR extraction, creating relevance labels, scraping, multilingual evaluation, custom model installation, fine-tuning, reranker comparisons, repeated chunk-size optimization, production vector-database setup, chatbot development or generated-answer evaluation. No guarantee that one model will win or that results generalize beyond the supplied corpus and queries.