Andrija Birtić

Front desks type fast, booking channels import blind, migrations copy everything. Across multiple property systems the guest base grew to 1.3 million profiles, and 26.3% of them belong to a person who already has another one. Histories fragment, loyalty misfires, the same guest gets the same offer three times.

A wrong merge poisons two guests' histories with each other's stays. A missed duplicate just repeats an offer. That is why the decision is three way, merge, reject, or ask a human.

There was no labeling budget, and none was needed. A proxy signal already sitting in the data turned out to be near ground truth for this problem, so the labels came for free, at corpus scale. What survives unlinked in the data is exactly what humans never caught. The model trains on the residue of the problem it solves.

Blocking is the search index itself, lexical and vector retrieval fused into one candidate list, no hand built blocking keys, near total pair completeness. For the decision itself, four resolvers competed on one shared benchmark, a weighted heuristic, a small feature based neural net, a cross encoder, an LLM judge.

Every resolver faced the same benchmark, held out clusters none of them had seen. The expensive favorites lost. The small feature based neural net reached 0.95 precision, trains in seconds on a CPU, and runs inference in milliseconds.

Nothing gets recommended while the sources disagree about who the guest is and what a stay even looks like. So the first thing built was agreement. A canonical data model for the whole domain. Kafka connectors pulling every property system into it. ksqlDB and a cleaner service standardizing events, quarantining the hopeless ones. And the deduplication layer on the stream, the deduplicator that earned its own case study.

No single description of a guest is enough. The model reads each guest through several views at once and learns to weigh them, leaning on whichever carries real signal for that particular guest. Build a guest below and watch the ranking follow:

Dedicated document scanners are expensive, lag years behind new document formats, and the rarer IDs and travel documents are often not supported at all. Every fix waits on the vendor.

So the scanner replaces them with whatever camera is present. Passports, national IDs, driving licences, and health cards, read at the desk and submitted as structured fields to every system that needs them. And training such a pipeline is hard for one reason above all, identity documents are sensitive personal data under GDPR. You cannot collect a training set of real ones.

Version one is a pipeline. Detect the document, warp it flat, fix the orientation, classify the type, find each field, read it, fuse the results. A model per stage, designed and trained in PyTorch, exported to ONNX, all running locally inside the desktop app.

The OCR was prototyped in house, a custom transformer OCR designed from scratch and trained on the synthetic lines, then benchmarked against a specialized classic engine. On par on accuracy, at a smaller memory footprint. A conv encoder reads the crop, a transformer decoder attends its way across it, character by character.

Version one earned its place, around 90 percent balanced accuracy on exact field match, and in production it reads nearly flawlessly. At around a second per document. But its pain lived at the edges. A document type it had never seen, a border pass, came through degraded, and every improvement meant retraining a whole family of models.

Version two replaces the entire extraction stage with one fine tuned vision language model. The desk app literally texts it, a document photo goes in, one strict JSON object comes back. It is even trained to refuse bad shots, glare or a covered corner gets a warning to adjust the document instead of a bad read. In practice it generalises far better on unseen niche formats, exactly where version one degraded.

Models from 0.6B to 8B were tried across vendors, the sweet spot settling around 2B and 4B, similar in speed and accuracy, served on a single consumer GPU at 96.6 percent field accuracy, under three seconds a document. Version two is in final fine tuning and early testing.

The same system ships two ways. A desktop app with local inference where the hardware allows it, remote inference where it does not. Either way there is one point of control, a central server pushing delta updates, so every desk runs the version it should.

Review dashboards give a business a star average and a list of frequent words, a score with no instructions. The reviews themselves are scattered across platforms, written in several languages, and polluted by namesake businesses on other continents. And an LLM asked to summarize them will happily paraphrase, invent, and silently skip. The problem is getting from that mess to a to do list you can trust.

One failure mode is precision and scope. Scrapers are unreliable, so the platform is built to know its scrapers and control them. Several scrapers per platform, so one failing means a fallback.

Search a hotel name and you get the right one plus its doubles on three continents. A layered identity filter kills the impostors, a review only enters the analysis when every layer agrees it belongs to the right business.

The other failure mode is the LLM itself, and it is tamed with orchestration rather than hope. The current generation runs as a LangGraph flow with structured output at every step, which turns a moody generalist into a robust, well performing worker.

An LLM will still sometimes skip reviews in a batch, short and foreign language ones especially. The loop notices and recovers, so every review gets an answer and none get invented.

Every claim the system makes must quote the review verbatim, an exact substring, checked. Every number in the report reconciles back to an extracted aspect, every aspect back to a quote, and every quote back to the original review on the platform it came from.

The report is category cards, source and month breakdowns, rising and declining keywords, critical issues with severity, and deduplicated action items each with a suggested fix.

725 modules of binary source, converted to 400 MB of machine generated markup that is mostly noise. And nothing stands alone. A module reads tables another one writes, calls shared libraries, inherits from common ancestors. The real problem is separating behavior from boilerplate and cutting one module free of the tangle, not translating syntax.

Every module goes through a full grammar for the legacy language into a Neo4j knowledge graph. Clustering the graph finds whatever repeats, screen shapes, trigger layouts, data access patterns, and the corpus collapses into a handful of families. Migrate one member well and the same recipe carries across its family, biggest families first.

The agent never reads source files. It queries the graph through a small set of narrow tools. The graph hands it the full closure of what one module needs, its screens, its triggers, the tables it touches, the libraries it calls, so a single module migrates standalone, keeping context lean by design.

The agent migrates a module, the result must build and pass its tests, then a human reviews and tests it for real. Findings feed the next cycle, and the cycles repeat until the migration is done.

Language models exhibit effective in-context adaptation, but this adaptation does not persist beyond the context window. Existing approaches address this by editing model weights. We instead ask whether a frozen model can retain knowledge acquired in context and reuse it across sessions. We attach a compact learned memory module to frozen open-weight models spanning 0.6B to 8B parameters, and provide a small number of in-context demonstrations of behaviors such as coding practices, formatting styles, and naming conventions. After the context and KV cache are fully cleared, the behaviors persist. In preliminary runs, demonstrated habits survive at or near ceiling while arbitrary associations such as specific names or values collapse to zero. Transplanting one session's memory state into a new session carries the rule to content neither session contained. In its current form, the memory generalizes within the family of demonstrated behaviors and does not yet carry novel rules outside it. Results are preliminary, and the mechanism, full quantitative results, and ablations will appear in the preprint once the remaining experiments are complete.

Software developer, tech lead and architect. I mainly specialise in building dense neural networks and the systems that use them end to end.

I care about honest evals, clean architectures, and solutions that actually make a change.