It’s been two years since we wound down Tactic. I’ve written plenty about what went wrong (the split focus, the premature scaling, the unit economics). But I’ve never written about what went right. And a lot went right.

Tactic ran for five years. It started as QuantCopy in 2019 and wound down in early 2024. We built a platform that helped sales teams research their target accounts by pulling and structuring information from anywhere on the public internet. At its peak we had a real-time web search engine, a natural language extraction pipeline (in the pre-LLM era), GPU clusters running our own models, and an auto-scaled Kubernetes infrastructure processing hundreds of gigabytes per hour.

I want to talk about what made it special.

We built something genuinely hard

Most B2B SaaS products are thin wrappers around a database. Tactic was not that. At its core was an Elasticsearch cluster wrapped in our own Index Management System, which let us run online migrations and A/B tests across multiple indexes with zero downtime. Feeding that was an ingestion pipeline that crawled the web (news articles, PDFs, annual reports, job postings) with selective OCR and table extraction.

On top of that sat our NLP extraction and summarization engine. A user could ask an arbitrary question about any company, like “What sustainability initiatives has this company announced in the last 12 months?”, and we would search the web in real time, retrieve candidate documents, re-rank them through an embedding system, and extract typed answers using a language model. The answers came back structured: numbers, dates, short text, paragraphs, picklists. You could sort, filter, and push this data straight into Salesforce.

This was not a lookup against a pre-built database. Every query hit the live internet. That’s what made it unique, and that’s what made it so technically demanding.

We were early to LLMs in production

When ChatGPT’s APIs launched in early 2023, we integrated them within weeks. We had spent years building our own in-house models for classification, extraction, and summarization, and suddenly there was a dramatically better option available through an API. Our team moved fast. We swapped out components where the cost-performance trade-off made sense and kept our own models where they didn’t.

The result was a step change in reliability. The thing that had held us back for years, the ability to capture relevant information from messy, unstructured web pages and return accurate, structured answers, suddenly worked. Our “custom data points” feature went from impressive-but-flaky to genuinely reliable. We were exploring hosting LLAMA 2 ourselves for tasks where it would be more cost-effective than OpenAI.

In retrospect, we were doing RAG before anyone called it RAG. Retrieve documents from the web, re-rank for relevance, extract answers with a language model, type and structure the output. That pipeline is now the foundation of half the AI startups that launched in 2024. We were running it in production in 2023.

The team shipped at an absurd pace

After our 2023 pivot to focus on CRM enrichment, we hit the highest feature velocity we’d ever achieved as a company. Multiple user-facing updates went out every week. Engineers got on calls with users and iterated based on what they heard. We didn’t have separate product managers filtering feedback through a ticketing system. Engineers talked to customers and shipped the next day.

I’ve written before about my philosophy on shipping fast: if someone gives you feedback and you act on it the same day, they’re much more likely to give you more feedback. That compounding loop was real at Tactic. Our best customers gave us incredible feedback because they knew we’d act on it immediately.

We also had a culture of making the smallest change necessary to test a hypothesis. I always respected engineers who could find a two-line hack to get something into production rather than planning a large refactor. That bias toward speed meant we could test ideas quickly and kill the ones that didn’t work before investing too much.

The people were exceptional

The age-old wisdom is that you become like the people you spend the most time with. I took that seriously, and Tactic was where I found some of the most talented people I’ve ever worked with.

I co-founded the company with Rudy Lai, who brought the original vision and a relentless commercial drive. Freddie Russo, Alex Sparrow, Augustas, Charles Pierse. These people pushed me to see what I was capable of and taught me what “great” actually looks like. They were productive, action-oriented, and held a high bar. Our team had deep expertise across data engineering, information retrieval, NLP, and machine learning, and everyone was expected to contribute beyond their speciality. Engineers owned deployment and observability. Designers shipped code. Everyone could empathise with the customer.

Working with these people was the single most valuable part of the experience. The skills were a close second.

I learned more than I thought possible

When I joined QuantCopy in 2019, I was a physics PhD who had taught himself Python by making graphs. By the time Tactic wound down, I had built and operated production systems across: distributed architecture, DevOps (infrastructure-as-code, CI/CD, monitoring, container orchestration), data engineering, cloud computing, NLP, generative AI, search, machine learning, and web crawling at terabyte scale.

Beyond engineering, I learned product design, first managing a design agency, then taking over design entirely. I learned product strategy, project management, roadmapping, alignment with customer success and sales. I navigated SOC II compliance, information security, vendor assessments. I hired, mentored, managed performance, and built culture. I fundraised, hired executives, and set company strategy.

There is no other experience in my career that could have given me this breadth. A PhD gives you depth in one narrow area. A big company gives you depth in one function. Being CTO of an early-stage startup gave me serious operating experience across all of it.

The vision was ahead of its time

When Rudy first pitched me QuantCopy in late 2018, he described monitoring the entire internet and using that information to trigger automated actions. It sounded like a path toward general artificial intelligence. I was fascinated.

That vision evolved over five years, but the core ambition never shrank. What we were really building was a tool that could summarise large amounts of unstructured text interactively: search the web, extract structured information, and make it immediately actionable. I wrote at the time that this had never been done before.

Two years later, this is essentially what every AI agent startup is trying to do. “Search the web, extract information, take action” is the basic architecture of Perplexity, of ChatGPT’s browsing mode, of dozens of AI sales tools that launched after we wound down. We were too early, under-resourced, and focused on the wrong market, but the core technical insight was right.

It shaped how I think about building

Tactic gave me strong opinions. I believe in shipping fast and incrementally. I believe engineers should talk to customers. I believe in the two-line hack over the premature abstraction. I believe you should eat your own dogfood, and if you can’t, that’s a red flag. I believe you shouldn’t scale your team past six people until you have real product-market fit. I believe in focusing on one persona and one value proposition until you’ve dominated that beachhead.

Some of these beliefs came from getting it right. Most came from getting it wrong. But they’re all grounded in real experience, not theory.

---

Tactic didn’t become a billion-dollar company. It wound down after five years with a quiet ending. But it was the most formative experience of my career. The technology was real, the team was exceptional, the learning was immense, and the vision turned out to be prescient. I wouldn’t trade those five years for anything.

---

Annotations: 0,8369 SHA-256 661bdc908ba932721c3eea19b9ee9de2
&Claude: 208,2096 2317,6052
@Jack: 2304,13
…