PragmaCraft

Ten years ago today I wiped a whiteboard and wrote: “KEEP CALM & LSTM!”

I was building our "Intelligent Search Assistant" project, pulling answers from search results with seq2seq and early NLP. It felt ambitious and fragile at the same time.

Here is a hard-won lesson learned in a decade: RAG isn’t a checkbox. The thing that makes or breaks it is evaluation. Real questions are messy and diverse. The valuable ones jump across documents, demand query rewriting, multi-hop reasoning, and tables/figure parsing. If you don’t measure that, you ship a demo.

We learned it the long way, first with our own trained reranker, now with advanced RAG, multi-step retrieval, iterative reasoning, strict guardrails, and traceable citations, all measured against ground truth and real failure modes.

How our stack evolved

- Models: from LSTMs/seq2seq to early transformers (BERT) to cross-encoder rerankers, kept our own reranker longer than fashionable because it worked on noisy enterprise text.

- Retrieval: from dense to query rewriting and multi-hop retrieval when the question demands it.

- Document understanding: from plain text and PDFs/OCR to reliable table & figure parsing, version tracking, and citations that land on the right cell/row.

- Evaluation: from smoke tests to ground-truth sets, adversarial cases, failure taxonomies, and release gates tied to business SLOs.

- Ops & governance: from lab boxes to on-prem, bank-grade deployments: RBAC, auditability, PII filters, latency budgets, air-gapped realities.

- Safety: from rules to guardrails that run within strict GPU/latency envelopes and are measurable (not vibes).

What hasn’t changed

- Vision: Treat knowledge as a strategic asset.
- Problem: Find truth, cite it, keep data safe.
- Principle: Shipping beats slides.
- Bet: Trustworthy on-prem GenAI will win in enterprises.

What’s next

Retrieval you can trust, reasoning you can measure, and assistants that make real work easier.

If you’re planning an on-prem RAG initiative or want an evaluation on your own data, let’s talk.

It's yo bert-day!

We're 7 years old. After seven awesome years full of bleeding edge Deep Learning & NLP research, we are excited as never before, to solve the most challenging problems, to create the best AI products, to provide real value for our beloved customers, thanks to our amazing team of brightest minds. Stay tuned!