LLM/agent systems that can be trusted
Evaluation, immutable feedback, calibration, and a human still controlling the action. Not an open text generator.
open something interesting
This page is better at answering curiosity than reciting a résumé.
earlier ink
local conversation card · editable · unsent
Nothing here is saved or sent. Copying is the only handoff.
about
I build the boring-but-important layer that makes LLM and agent systems safe enough to operate: evaluation, feedback, provenance, and authority boundaries.
If a system is wrong in ways that are hard to see, I want to make the failure visible and fixable before a customer meets it.
- experience
- 10+ years · production systems · backend · data
- now
- LLM/agent evaluation · governed automation · human-in-the-loop
- also
- native Android and offline-first when the problem needs it
- looking_for
- a real operational or data problem, not another dashboard
- mobility
- open to relocation
focus
Production reliability
Observable failures, provenance, and controlled writes. Make the system explain why it did what it did.
Android and local-first
Native apps that keep ownership on the device and respect privacy. A second lane, not the whole story.
experience
Badoo · Ozon · Bits4Sky
More than ten years in production software
- systems
- monitoring · backend engineering · automation · data platforms
- today
- LLM/agent evaluation · human-in-the-loop automation · native Android
- through-line
- make hidden failures visible · keep humans in control · ship the boring-but-important part
The titles changed; the useful part did not: trace what is actually happening, understand the whole system, build the missing software, and make it explainable to the person operating it.
Full experience on LinkedIn Part-time Physics student · The Open University
projects
Review-response LLM evaluation systemproduction · anonymized
Production LLM workflow with immutable feedback provenance, model evaluation, calibration, and human-gated writes.
backend · production · LLM evaluation · feedback loop · human-in-the-loop
agent playground · investigate a suspiciously simple question
Agent-governance harnesspublic source · sanitized
A persistent control loop that turns a real goal into observable claims, dispatches bounded agents, and stops for a human at real authority boundaries.
backend · production · reliability · automation · human-in-the-loop
VoiceGuidesworking pre-release
Native Android listening for local guides: continuity, anchored notes, and offline ownership.
android · offline-first · reliability · privacy
Earnedactive development
A local Android experiment in useful feedback without coercive gamification.
android · offline-first · human-in-the-loop
EcoWatchprototype
Offline-first field capture for intermittent networks and imperfect coordinates.
android · data · reliability · offline-first
contact
Already know why you’re here? Borrow an opener:
for a role
“Hi Valia — I’m hiring for [role]. The interesting problem is [one sentence]. Would you be open to a conversation?”
for a system behaving mysteriously
“Hi Valia — reality and our dashboard have stopped agreeing. Want to investigate both?”
for an early idea
“Hi Valia — we have a useful nuisance that may deserve software. Can I tell you about it?”
open coast.jpg
