I am currently serving as a technical lead at a Finland-based startup: Linkenite Oy. We develop end-to-end automation pipelines for our clients and also develop intelligent systems to solve real business case studies. Now for the topic of the session on day two of AI DevCon: Turning generative artificial intelligence (GenAI) into business value.
Nowadays, anyone can get access to application programming interface (API) keys for any foundational large language model (LLM), and it is very easy to build your next GenAI system.
Monetising GenAI to generate business value
The most important aspect is that you build a system that can be relied on and, most importantly, that can be converted into business value. And that is what we are going to go over. We built an end-to-end pipeline for the composites industry with the help of a deep 4-layer architecture. Before moving further, I would like to ask: How many of your GenAI systems are actually generating measurable business return on investment (ROI)?
We had a Europe-based client who has 20-plus years of experience in this industry, working at the intersection of three main domains: investment, composites and finance. Now, this client of ours had a major career shift from a full-time job to being a solo entrepreneur.
Time was the most valuable resource that this person had. Everyone wanted his decades of experience, but even a single mistake showing up in his newsletter or in his articles would be a catastrophic loss of his credibility because everyone looked up to him. This individual was facing a major challenge where he needed to build a system that collated the authentic information of this niche-based industry as well as mimicked his exact brand DNA.
Most importantly, he wanted a system that never lied and that never hallucinated and that is where we came in. We built an enterprise intelligence system for end-to-end newsletter automation.
Measurable impact
This newsletter generated (after six months) around 2000-plus active readers, 100-plus paid subscribers and 50-plus paid consultation calls; overall, a 28 per cent productivity gain. Also, there was a reduction in the research time from three hours to 15 minutes. We ran this automation for 15 minutes and we got all of the news along with the draft newsletter for our client.
The nitty-gritty
It was very important to generate real revenue as soon as possible without risking his credibility and losing any business. We developed an architecture: a 4-layer strong architecture that focused on:
- Autonomous discovery and ingestion
- Context-aware ranking
- Stylistic retrieval-augmented generation (RAG) and memory
- Forensic audit
Building a system that was trustworthy and reliable was the need of the hour; also, we needed to mirror his expertise.
Deep dive
In discovery and semantic ranking, we know that the news in the information era is spread across the internet. So, what’s important is to find the authentic information that is predominantly present behind data-protective ecosystems such as LinkedIn, PDF archives, authwalls and paywalls. We solved this problem using a tool called Playwright.
We built a headless orchestration layer using Playwright because LinkedIn is a data-protective platform, but authentic news and information can be found on the same. We used Playwright along with a function that is a context-persistent session. The session of the user is protected by storing the machine configurations, the digital fingerprints of the machine and the session cookies. Hence, your internet protocol (IP) address is not detected by the machine learning algorithm of LinkedIn, and is not flagged and banned permanently from the LinkedIn platform.
Now, in this discovery of proper information and proper news for our client, we did solve another sub-problem, which was keyword heuristics: a heuristic path that we ran through the database, which looked for similar keywords.
We created a bag of 2000 words including keywords and terms like ‘series A’, ‘series B’, ‘mergers and acquisitions’, ‘patent’, ‘innovation’, ‘investment’, ‘funding’ and so on.
In stage two, we performed LLM re-ranking. We passed it to the Perplexity Sonar Pro model, which created a relevance score against a criterion: that any particular news item or data point should have a maximum relevance score. That is to say, it has to match the maximum number of keywords from the bag of words.
So, the more keywords matched, the higher the relevance score. That is how we could rank it against the LLM. Now, why did we use this particular model? The reason is that the F-score of Perplexity Sonar Pro is 0.86 (the F-score being the benchmarking criterion for all LLMs). Thus, Perplexity Sonar Pro is highly ranked for any research-driven work.
So, stage one was keyword heuristics, wherein we passed the keywords and performed keyword matching. Stage two was LLM re-ranking, where we ranked all of our data based on the innovation or relevance score criteria.
That is how, in stage one, we ensured that we not only discovered the right news by passing the authwalls, paywalls and data-protective ecosystems but, at the same time, we also ensured that we were only passing on high-signal data that moved ahead in the pipeline.
Coming to stage three: stylistic RAG (pattern mimicry). Generally, RAG is used in order to augment and generate the data. Retrieving implies data retrieval, augmentation entails maximising the data and then the data is generated.
Now, we needed to mimic our client’s brand tone and his persona. So, we did use RAG for something different. We used it for pattern mimicry. We used it to make sure that the tone and the style in which this newsletter would be written matched the precision of our client and his brand DNA.
For that, we selected the Facebook AI Semantic Search (FAISS) database for historical data and newsletters. So, we took the past two years of data from our client and we ingested this into the FAISS database. It is Facebook’s open source database that is often used for LLM database storage and vector search.
We performed the extraction of structural layout markers and analysis of investor-grade tonal intensity to evaluate which news is important for our clients.
The mindset with which our client selected the news and marked a piece of news as really important, or brought which news into which section of the newsletter, was also vital and subjective. The taxonomy, the tone – everything mattered.
For this, we created a style JavaScript Object Notation (JSON) data for formatting, tone and taxonomy. This style guide JSON was then passed on to our LLM with strict four categories. The formatting was bold company name, tone was analytic implications and taxonomy was a strict 4-category map.
Now, all of this was being derived from the historical data and newsletters that we had picked up from our client and we did perform stylistic RAG on it to identify his exact tone and pattern.
On to stage four. For any given GenAI that we build, it is very important for us to run it through a forensic audit strategy. We basically created a two-stage framework.
Stage one was atomic claim extraction. Here, we passed the entire draft of the newsletter to this LLM, which then broke it down into a list of verifiable statements. Say we have a huge newsletter draft that contains somewhere around 3000-5000 words. Now, those 5000 words were divided by this LLM into a list of verifiable statements. These statements (containing numeric claims, dates, amounts, entities, company names) and all of this data were stored together in a ‘claims.txt’ file.
Extraction first involved passing the entire draft and breaking it down into verifiable statements with all of the precise data points. In step two, we again claimed whether it was true or not. A value of zero meant that the data was false and one meant that the data was true. If the data was true, then we passed it on in the final newsletter draft. Otherwise, we scrapped it and eliminated it from the newsletter. In the last stage, it was curated into ‘validation_report.txt’. So, after claims.txt, we moved to validation_report.txt.
The ‘Custodian of Truth’
For any production-grade system, I would like to emphasise the fact that no matter how authentic or powerful our GenAI system is, we should always choose the human in the loop to be the custodian of truth. A human would understand any discrepancies that the AI might have incorporated in the newsletter draft.
This was the business ROI that we created for our client. Autonomous infrastructure compressed the manual research time for the newsletter, which was not written by a human but generated by an autonomous end-to-end newsletter engine and pipeline.
Snippets and recap
Okay, I have a couple of questions from the case study. Let us answer them one by one.
What does the ‘trust hurdle’ refer to in a specialised industry like advanced materials? It is the requirement for surgical accuracy where hallucinations and financial data become significant liabilities.
What does the expression ‘Tokens not equal to revenue’, in an industrial GenAI context, imply? That generating high volumes of content does not automatically translate into measurable business value.
To recap, the four stages of GenAI orchestration architecture were autonomous discovery and ingestion, RAG for pattern mimicry and tone identification, then memory and the last stage was forensic audit.
In the ingestion phase, the ‘launch_persistent_context’ helped us to scrape the data and bypass the protective machine learning algorithms that LinkedIn, as a data-protective platform, implemented. We could run our automated scripts with Playwright and the built-in function.
Stage two (ranking) employed a two-stage filter where stage one was about LLM extraction and the next stage was about re-ranking.
In the discovery and semantic ranking process, the function of stage one keyword heuristics was scoring the posts based on terms like ‘breakthrough’, ‘scale-up’ and ‘mergers and acquisitions’.
In stage two of the ranking process, to evaluate raw scrapes against innovation relevance criteria, we used Perplexity Sonar Pro as it is apt for any kind of research work.
In the context of the provided architecture, vector stores were used for storing facts, encoding and mimicking brand tone.
The dynamic style in the context of stylistic RAG is a set of patterns that were extracted from historical content regarding headers, bolding, and tonal intensity.
You can use one LLM to solve multiple problems. It is all about experimenting and getting the right things done.
To mimic the brand tone and understand the taxonomy precisely with respect to a niche industry, a taxonomy map was created.
So, what was the chunk size used by the ‘RecursiveCharacterTextSplitter’ in the stylistic RAG pipeline? It splits the entire text – the data that we provide to the database – into smaller chunks. Each chunk was determined by the number of tokens it consumes. So, it had around 1000 characters per chunk.
Also, the primary function of the forensic audit layer was to provide an automated fact-checking pass to ensure the AI output contained no inaccuracies.
Perplexity Sonar Pro API performed real-time web verification of claims in the forensic audit stage.
The document generated to flag discrepancies (such as currency errors) for human review was ‘validation_report.txt’.
In the forensic validation loop, the designated title for the human is the ‘custodian of truth’ – the human in the loop who verifies every statement.
In the hybrid system architecture, which tool served as the hands? We had a mixed architecture. General scraping Python scripts would be banned in any data-protective ecosystem. We used Playwright. This is a very powerful tool for any automation task that you need to perform, such as scraping data that is hidden behind paywalls.
FAISS served as the memory in the hybrid system architecture.
What percentage of the work in production-grade AI should be dedicated to validation? 50 per cent. So, 50 per cent of the work should be done by your automation workflows, but the remaining 50 per cent should not be done manually. The human in the loop should just cross-verify it, but 50 per cent of verification should also be performed with GenAI, with the help of the logic that you build in.
Here, we generated the draft with the help of an LLM and we then used another LLM, which is good at research, to verify the same draft. So, did not involve any human effort and at the same time we were not passing any random facts that would destroy the credibility and trust of our system.
Under what conditions does the orchestrator script skip the discovery phase? If the file context.json already exists. If you already have the proper context, it is valid to skip the discovery phase.
Now, what was the specific inaccuracy found in the Series B funding claim? The actual funding was in US$. This is a huge mistake that an AI or LLM can make. So, for our client, if funding is being performed in US$ but the system incorrectly reported it in Euros, it would be a major difference in the facts and the numbers. And that is something which you could track at layer four of this architecture, that is forensic auditing, when this particular inaccuracy would be spotted and either corrected at that moment or rejected from the newsletter.
The measurement of productivity in any production-grade system is done on two parameters. One is the metric of how many hours of human work are being replaced by the automated engine. The time consumed was three hours. Now, our automated system can compress that time into 15 minutes. So, one metric is time. Another metric is accuracy.
How we start building any automation engine at Linkenite is that we take into account the manual workflow – how the client used to perform it. We took all the manual steps that our client performed, broke them down and replaced them with the right tools like LLMs, and the right techniques like RAG.
This article is based on the talk, ‘Turning GenAI into Business Value – A Practical Case Study’ by Tisha Singh, Linkenite Oy, at AI Dev Con India 2026. It’s been transcribed by Anirudh Shobhanan, Electronics For You.
