Production systems we designed and operated

Sitting between bot services and local LLM backends, it processes each request through a pipeline of classify → run role agents in parallel → score confidence → merge. Memory and retrieval (RAG) stay the responsibility of the calling service, so the gateway focuses purely on orchestration and safety — a deliberate boundary that keeps domain logic out of the gateway.

What sets it apart

• Adaptive routing Automatically distinguishes empathy-oriented from analytical/reasoning responses, switching to a hybrid path that runs both models in parallel only when needed.
• Cost-aware merge gating When two candidate responses are confident or similar enough, it skips the extra synthesis LLM call — cutting unnecessary compute on a single-GPU setup.
• Fault-isolated resilience Circuit breakers and health probes let it fall back instantly to a live model when a backend goes down, without blocking on timeouts.
• Real-time token streaming Single-model paths stream backend tokens straight through for responsiveness; hybrid paths send after merging to preserve completeness.
• Layered DB guardrails When executing LLM-generated queries, read-only access, syntax validation, and target-table restrictions are stacked to structurally prevent sensitive-data exposure.
• Zero-downtime config reload Model endpoints and policies re-apply automatically on file change, so operators tune behavior without restarts.

A case study in safely exposing the web surface of a service that handles sensitive personal data. Sensitive functions such as coaching and editing stay on the messenger bot, while the public web is restricted to view-only to minimize the attack surface. The pages run with no separate build step — a deliberately lightweight stack.

What sets it apart

• Read-only privilege separation The public-facing path holds no write permission at the database level, so even an application flaw cannot alter data.
• Token-leak prevention Authentication tokens are kept out of server access logs, and each page runs under a least-privilege content security policy.
• Stateless authentication Signed ephemeral tokens handle auth with no session store, staying robust across restarts and multiple instances.
• Server-rendered visualization Visualizations like relationship networks are rendered as SVG on the server, delivering rich views with no client-side scripting.
• Dual mobile/PC experience Supports both the in-app messenger (WebApp) and regular PC browsers (plain HTML forms), each under its own security policy.

The essence of this project is not "a bot that earns well" but the engineering that keeps a flawed strategy from leaking into live trading. It is built on the premise that naive backtests easily manufacture false edges, and that trusting them unverified is the biggest risk. So both research-stage discipline and execution-stage safety gates are enforced at the system level.

What sets it apart

• Backtest-illusion discipline Survivorship-bias removal, conservative fill assumptions, in-sample/out-of-sample and monthly decomposition, and an alpha-over-benchmark criterion are baked into the validation process to filter out false edges.
• Single source of decision (backtest ↔ live) Universe, entry, and exit logic are shared by the same code in backtest and live execution, structurally preventing drift between the validated strategy and actual behavior.
• Three-stage promotion (shadow → paper → live) A strategy moves through simulated then forward-paper validation, and only promotes to live after meeting realized-performance criteria and human approval.
• Fail-closed safety gates A live order only goes out when all safety conditions hold; if any one fails, the system automatically closes to "do not order" — a safety-first design.
• Per-market process isolation KR and US engines run as independent processes, with a lock preventing duplicate runs on the same account to avoid order conflicts.
• Natural-language operations control Operators can pause or adjust the engine via messenger using natural language and commands, while risky actions require two-step human approval.
• LLM-debate strategy research An autonomous loop generates, backtests, and rotates strategy candidates via LLM debate — still behind a human approval gate.

※ This item is a technical reference to showcase software engineering capability. It is not investment advice or a solicitation to trade, and guarantees no returns.