product-data enrichment
We build AI systems that survive production.
BARGO is a two-founder data R&D house in Rīga. We take AI from demo to production — agentic workflows, RAG pipelines, model evaluation, AI integrated with the systems your business already runs.
We ground every result: each value carries the source it came from, or it ships blank. The proof is live: unbin, our product-data enrichment engine at unbin.io — and a published record of what we deleted when the measurements said no.
Grounded by default. Nothing is invented.
[ I ] BARGO · RĪGA SIA · REG 42403048572 · SINCE 2020
[ II ] FOUNDER-LED EVERY FIELD SOURCED
[ III ] WHAT WE DO PRODUCTION AI SYSTEMS · GROUNDED DATA · EU-REGULATION RESEARCH
[ IV ] CAPACITY 4 ENGAGEMENTS / YEAR · NEXT: Q4 2026
[ 01 / THE HOUSE ]
What we build.
Three entries — one live product, one core research service, one published source repository. Each carries the same discipline: every field cites its source, or it ships blank.
EU-regulation research
core answered from the primary source
published source
[ 02 / SEVEN PRIMITIVES ]
Seven primitives from the live runtime. Five run by default; two are held off until the numbers clear.
Seven primitives sit in the runtime that powers unbin.io. Five run by default in the live engine. Two are held behind flags by their own measurements: the injection fence is built and unit-tested, but its first paired A/B regressed extraction quality, so it stays off while the guard clause is reworked and re-tested; the active-inference fetcher's stop policy was refuted on a 91-SKU A/B, reworked, and stays off until the full benchmark re-clears it. Each primitive below traces to a file and is worth reading about on its own. If your team is fighting hallucinations, ungrounded confidence scores, prompt injection, or pipelines that fail silently — these are the mechanisms we bring to an engagement. Six of the seven are published as readable source, with tests, at github.com/bargo-lv/primitives.
Per-field fusion. runs by default Resolves conflicts across multiple sources for the same field, weighted by learned source reputation — TruthFinder / Accu-style, with a Beta-Bernoulli prior that updates as each source proves or disproves itself over time.
Calibration monitor. runs by default Keeps confidence scores honest with isotonic + Platt calibration and an ECE drift alarm that fires the moment stated confidence stops matching observed accuracy.
Self-verify. runs by default Checks every cited scalar against the evidence the run read from its cited URL — a value its own source contradicts is nulled before the field ships.
Determinism alarm. runs by default Per-organisation completeness-score regression detector plus a silent-seam liveness probe — a pipeline that quietly stops producing evidence gets flagged instead of failing quiet.
Active-inference fetcher — refuted, then reworked. off · pending full benchmark A ranked-fetch policy grounded in Friston / Weitzman break-even reasoning. An A/B on 91 SKUs cut cost by 76% but regressed grounding by 40% — the 8-SKU pilot had looked clean; the full run refuted it. We reworked the stop policy and keep the mechanism off by default until the full benchmark re-clears it. Publishing the refutation is the point.
Injection fence. off · pending paired A/B A deterministic prompt-injection envelope — untrusted source text passes into the model wrapped and defanged, so it cannot rewrite the instruction it was meant to inform. Built and unit-tested; off by default until a reworked variant wins its A/B.
ETIM 10.0. licensed data · in every enrichment Licensed classification data — 5,640 classes, 76,625 feature mappings — modelled internally for enrichment. (ETIM International, Open Data Commons Attribution Licence v1.0.)
We also deleted an entire product-carbon-footprint capability from the same runtime. The receipt is in the post-mortem below.
[ 03 / POST-MORTEM ]
What we deleted while shipping. What we chose not to build.
The primitives above come from unbin, our product-data enrichment engine live at unbin.io. Along the way we A/B-tested mechanisms inside the same runtime and logged the result whichever way it landed. Three receipts carry more explanatory weight than the primitives combined.
9141fbd5 cost −76% · grounding −40% reworked · off by default A 91-SKU A/B refuted our active-inference stop policy. The earlier 8-SKU pilot had looked clean; the full run caught what small samples hide.
0063afe3 no safe variant at n = 44 deleted outright Paired A/B of a tool-allowlist mechanism: mean cost rose 17.8%, every grounding metric tied or worse. The code reverted to its original byte-identical form.
0bbe89cc 35 billings vs 719 for enrichment scoped out The product-carbon-footprint capability — roughly 5,950 lines and 17,446 emission-factor rows — removed once the billing record showed it was not what customers were paying for.
We ran the same rigour on our own product-shape question — should the engine scale beyond unbin as a distributor SaaS? Four candidate shapes, scored across six constraint axes, each judged through a default lens and a deliberately hostile contrarian one. None of the four cleared the constraint gate. unbin stays at its current shape; the engine underneath keeps being maintained rather than expanded. The scoring is published too.
This is our default posture. We ship, we run measured experiments in our own runtime, and when a mechanism doesn't earn its place under an A/B it doesn't ship — we delete it or hold it off, and we log the reason. When we score a scaling proposal and it doesn't clear our own gates, we don't build it — and we log the reason. The commits and the gates are the evidence.
[ 04 / WHO WE ARE ]
Two founders. One internal stack. Grounded by default.
BARGO, SIA has been operating in Rīga since 2020. Two founders. Every field sourced.
We run an internal substrate we don't disclose. What we disclose is what it produces — the primitives above, the deletions we made when tests failed, the post-mortems we write instead of press releases.
Everything rests on one principle: nothing is invented. Every claim and every number traces back to the source it came from — so anyone can open it, re-read it, and verify it. That discipline is the whole company.
Arnis Geidmanis
Founder & Architect
Twenty years of enterprise IT before BARGO: nine consulting on systems for Volvo Trucks, eight in the Würth Group's Baltik Elektro as ERP administrator, IT project manager, then head of IT — Microsoft Dynamics AX, business intelligence, multi-country delivery. Built unbin's engine end to end: crawling, grounding, evaluation, production operations. The architect on every engagement.
Gunārs Auziņš
Co-founder & Strategic Advisor
Reads the regulation before it trends. Tracks CSRD, CSDDD, ESPR, CBAM and the rest of the EU alphabet to their primary sources, maps each to what we build, and keeps every claim we publish anchored to the text it came from.
[ 05 / WORKING WITH US ]
Four engagements a year. Fixed scope first.
We take a small number of engagements where production matters more than demos: AI agents and agentic workflows that run real business processes; RAG pipelines over your own documents and data; AI integrated with the ERP, CRM and internal APIs you already run; evaluation, observability and cost control for LLM systems already in flight. We also answer EU-regulation questions from the primary source. If a "production AI engineer" vacancy has been sitting open on your careers page, this is that capability — without the hire. Pilots are where those hires quietly fail: our own 8-SKU pilot looked clean, and the full run refuted it. We build for the full run.
[ I ] Scope
A written fixed-scope proposal: the deliverable, its boundary, its price — typically a day or two from a funded conversation. You know exactly what you get before you commit to anything.
[ II ] Build
The architect does the work — no handoffs, no juniors. Grounded by default: every claim and every value carries its source, or it ships blank.
[ III ] Continue
Optional hourly continuation if you want more. No retainer lock-in. The deliverable is yours either way.
The method is public: mechanisms that fail their A/B don't ship, and the receipts are published — the last one cut fetch cost by 76% while grounding fell 40%, and it still didn't ship. The machinery that runs it stays private, and so does your engagement. Capacity is four engagements a year; the next opening is Q4 2026.
[ 06 / CONTACT ]
Have a funded AI problem? Write.
Production RAG, agentic workflows, AI that has to live inside the systems you already run, evaluation and cost control, an EU-regulation question that needs the primary source — write to the architect directly. The first reply comes with questions, not a deck. EU-regulation question rather than an AI build? That inbox is gunars@bargo.lv.
No form, no funnel, no sales team. Two founders, two inboxes. Peers and collaborators are always welcome. Fixed scope first — the deliverable is yours either way.
Engagements: arnis@bargo.lv
·
Research & notes: gunars@bargo.lv