Ilya Galendukhin — Python · Go · PHP/Laravel

Backend, ML, & the security in between.

Backend & ML engineer.
I build matching, retrieval and agents
on systems that have to hold.
STATUS: LLM red-teaming · PortSwigger · CTF

PLATE_01 · SESSION · ~/about

Matching, retrieval, agents, and the socket underneath — the layer that has to be right.

Selected work · [ 04 ]

Selected work

01 // record linkage · LLM agent

Resume
Deduplication
for an ATS

Two-stage entity matching that refuses to pay the O(n²) tax. Cheap hard filters and blocking keys — phonetic, career-fingerprint and HNSW vector blocks — collapse the pairwise space before weighted fuzzy scoring ever decides a merge. I also built the candidate-matching funnel: applicants ranked against a role through cascading stages — hard filters → embeddings → reranker → AI-agent scoring — each stage cheaper and broader than the one after it. The canonical record is non-destructive and the scoring algorithm is versioned, so full source history survives and every result can be reprocessed as the model gets better.

  • Python
  • PostgreSQL
  • pgvector
  • HNSW
  • reranker
  • LLM agent

02 // fine-tune · function calling

On-Device
LLM Assistant

Fine-tuned Gemma 270M to actually call tools. I built the training data myself — 42,000 examples across six function types — trained on a single A100, and landed 76% function-call accuracy on the held-out set. The point was the constraint: a model small enough to run on CPU or a phone in roughly 1 GB of RAM — capability that doesn't need a datacenter to answer.

  • Gemma 270M
  • PyTorch
  • fine-tuning
  • function calling
  • A100
PLATE_02 · TRAIN.LOG · gemma-270m-fc

03 // LangGraph · SSH orchestration

deploy-agent

PLATE_03 · DEPLOY.LOG · langgraph

A conversational agent that takes a deployment from zero to running. It provisions a fresh server over SSH, reads a repository to infer its stack and which databases it needs, generates the Docker and CI config that was missing, and triggers the pipeline — all in one conversation. It turns the routine, error-prone parts of shipping into something you describe rather than hand-assemble.

  • LangGraph
  • Paramiko
  • GitLab API
  • SSH
  • Docker
PLATE_04 · HANDSHAKE · rfc1928/1929

04 // RFC 1928/1929 · networking

SOCKS5
Proxy in Go

The protocol implemented from scratch, straight off RFC 1928/1929 — method negotiation, the username/password handshake, address parsing across all three ATYPs, and full bidirectional relay. Not a wrapper around someone else's library; written for the networking depth, to know exactly what crosses the socket.

  • Go
  • RFC 1928/1929
  • networking
  • from scratch

Day to day · in production

Daily routine

Beyond the highlights, this is the work that fills most weeks — shipping a Go platform, owning its infrastructure, keeping a fleet of PHP services healthy, and tightening queries on every side of the database.

  • › Go · smart-bid + tracking

    My main codebase. I ship features into the smart-bid and tracking platform constantly — Go over MongoDB on the operational path and ClickHouse for analytics — and I write the infrastructure that runs it, not just the application code.

  • › Infrastructure & DevOps

    I own the path to production: GitLab CI/CD, Linux, Docker and Kubernetes. The infra is something I build — pipelines, images, orchestration — not just something I consume.

  • › PHP · Laravel — 5 services

    I keep five PHP/Laravel tracking services running, and I ran the Laravel 10 → 11 upgrade across them myself.

  • › Query optimization

    Constant work on both sides of the database — tightening the OLTP queries on the hot path and shaping the analytical queries alongside the analysts, so the numbers come back fast.

  • › Mentoring & code review

    I mentor two junior engineers on the team and review their pull requests — the slow, compounding work of turning 'it runs' into code that holds.

Manifest · what I reach for

Stack

Languages
Go · Python · PHP / Laravel
ML / LLM
PyTorch · LangGraph · fine-tuning · function calling
Data
PostgreSQL · ClickHouse · MongoDB · pgvector · HNSW
Infra
Docker · Kubernetes · GitLab CI/CD · Linux · Yandex Cloud · AWS
Security
LLM red-teaming · PortSwigger labs · CTF

Ping me.