Anton Koshcheev

Koloq.co·Applied AI Engineer (solo side project)·koloq.co·2025-05 — Present

• – Built a production LLM pipeline turning multilingual Telegram chat messages into structured rental listings (30+ typed fields: price, dates, geo, amenities). Prompts, validators and mappers are generated from a single tag definition — new fields propagate with zero manual edits. 7k+ listings extracted to date; 200–700 live listings in each major city.
• – Designed an internal LLM gateway so every model call passes one governed choke point: multi-provider routing (Vertex AI Gemini, OpenAI-compatible), per-route RPM/RPD/TPM quotas, a priority queue so batch backfills never starve interactive traffic, circuit breaking, retry classification (upstream 429 retries, internal quota doesn't), and OpenTelemetry/Grafana observability with alerting.
• – Cut classification cost ~40% with a 4-stage cascade (zero-cost heuristics → embedding classifier → batched cheap LLM → full extraction); every stage fail-open, rolled out per-source in shadow mode.
• – Trained the embedding relevance classifier (768-dim embeddings + logistic regression, 5-fold CV): F1 0.971, 0% cross-validated false negatives. Shadow evaluation caught a production domain shift (FN spiked to 13.6%); retrained on 8.1k examples with class-imbalance handling and threshold recalibration — FN restored to 0.7%, hot-swap deploy.
• – Built an eval harness over a hand-labeled golden set (150 examples, 9 strata, 20% held out): per-field accuracy with asymmetric FN/FP costs, JSON validity, p50/p95 latency, $/1k messages — with fabrication rate, not accuracy, as the deciding metric (cheap models don't fail loudly, they invent dates). Benchmarked 10 models against the production baseline and migrated to one matching quality at ~2.5× lower cost, with a documented rollback path. Every inference carries a SHA-256 hash of its effective prompt, so eval runs and production rows attribute to an exact prompt revision.
• – Rebuilt the product around behavior that already exists: killed the original form-based platform (profiles, messenger, listing forms) and relaunched as parsing of what people already post. The housing bet runs under a public 12-week stop-loss (10 signed-in users or the push stops); the vertical-agnostic extraction + map engine keeps running as a working tool.

TypeScript · Node.js · Python · scikit-learn · Vertex AI (Gemini) · OpenAI API · PostgreSQL/PostGIS · OpenTelemetry · Grafana

Kraahl·Founding Engineer·Hybrid, Paris·2025-10 — Present

• – Grew a collectibles auction marketplace to six-figure monthly GMV in 9 months — real-time auction engine, multi-provider checkout (Stripe, PayPal) across 6 currencies, cross-border shipping (Shippo), 220+ sellers onboarded.
• – Re-architected the founder’s LLM-prototyped (vibe-coded) features into a contract-first hexagonal core — one shared Zod contract package consumed by web and mobile — with a skill library and hard review gates so LLM coding agents produce architecture-conformant code by default.
• – Enforced fintech-grade testing on the money path: walking-skeleton pipeline tests for every money flow on both payment rails, 60 CI fitness tests that fail the build on architectural violations, and a real-database integration gate on every PR (173/173 green).
• – Shipped two production LLM features: an ops copilot — a tool-calling agent over the marketplace admin APIs (orders, refunds, partial payments, disputes), the operator asks “what happened with order 3154” instead of running probe scripts, every run trace-logged in Axiom — and multimodal (text + image) listing moderation with a persisted audit trail and human override.
• – Designed the transactional core (PostgreSQL, strong consistency under concurrent writes) and a self-healing money path — idempotent webhook handling, payment-recovery replay, 15-minute provider↔DB reconciliation — coordinating money across Stripe, PayPal and Shippo; zero data-integrity incidents in production.
• – Ran the product loop, not just the code: every feature shipped as a hypothesis against a driver tree, demand-funnel instrumentation to measure it, and in-database analytics (cohort retention, sell-through, seller-concentration risk) deciding what to build next.
• – Kept the site fast and up through drop-day traffic spikes: optimized hot paths for surge concurrency and rebuilt the image pipeline around an explicit CDN cache-key cost model, sharply cutting page weight and load times.

TypeScript · Node.js · React/Next.js · Expo/React Native · PostgreSQL · Supabase · Stripe · PayPal · Shippo · Axiom

Starknet Ecosystem·Open-source Engineer, OnlyDust grants·Remote·2023-05 — 2024-12

• – Grant-funded via OnlyDust: ~20 PRs merged across Beerus (Rust, Starknet light client), starknet-devnet (Python, testing infra) and Pathfinder full node — shipped in pre-spec territory, reverse-engineering undocumented hashing and encoding conventions across ~6 repos.

Rust · Cairo · Python · TypeScript

Ozon Tech·Backend Engineer, Retail·Moscow·2022-05 — 2023-01

• – Owned 3 gRPC services in the retail domain (~10 interconnected services, peak ~50k RPS, sub-100ms p99); built Kafka pipelines for the order-status flow.
• – Introduced Jaeger tracing and structured logging, cutting median cross-service debug time from ~3h to ~40min; led ~6 production incident investigations.

Go · gRPC · PostgreSQL · Kafka · Kubernetes · Prometheus · Grafana · Jaeger · TypeScript · React

Skyrex·Backend Engineer (Crypto Trading)·Moscow·2020-06 — 2022-01

• – Built the trading MVP from scratch: order-execution engine, portfolio tracking, Binance API integration, event-driven RabbitMQ pipeline (p99 routing < 150ms).