ML-powered item build recommendations for Valve’s Deadlock.
DataLock is a context-aware item recommender for Valve’s Deadlock. Select your hero, allies, enemies, and game phase to receive ranked item suggestions predicted to improve your chances of winning. Instead of static build guides, DataLock adapts recommendations to the lobby you are actually in, using an ML model trained on large-scale community match data.
Recommendations are enriched with item metadata and shop images, then presented in a clean, fast UI that makes it easy to compare options and lock in a build.
DataLock is split into three layers: a fast UI for selecting match context, a serverless inference API for ranking items, and a private model bundle that keeps the ML artifacts and metadata secure.
1) UI collects match context (Vercel, Next.js)
You select your hero, allies, enemies, and a game phase (early, mid, late). The UI is intentionally lightweight so the interaction feels instant and the recommendation flow is easy to understand.
2) Requests are routed server-side to protect secrets (Next.js API routes)
The browser never talks to AWS directly. Instead, the Next.js server routes forward requests to the inference API and attach an x-api-key from environment variables. This keeps the API key out of client-side code and prevents it from being scraped from network requests.
3) Serverless inference runs on AWS Lambda (Python)
The AWS HTTP API triggers a Python Lambda handler which:
- validates the request payload (hero IDs, phase, team sizes)
- transforms the match context into a fixed-order feature vector (using versioned metadata)
- scores candidate items using an XGBoost model trained on large-scale match outcomes
- returns a ranked list of items with scores and item metadata for display
4) Model assets are loaded from private S3 and cached for speed
The inference bundle (models + feature ordering + item/hero metadata) lives in a private S3 bucket. On cold start, Lambda downloads the bundle once, then caches it in /tmp and process-level globals so subsequent invocations reuse the assets without repeated S3 fetches. This reduces latency and cost while keeping the artifacts private.
5) Results are rendered as a build shortlist (Next.js UI)
The frontend renders the ranked recommendations as item cards using the returned metadata (names and shop images). The goal is to make the output feel like a usable build shortlist rather than a raw model score dump.
flowchart TD
%% =========================
%% DataLock High Level System
%% =========================
%% Frontend (Next.js on Vercel)
subgraph "Frontend (Next.js 16 on Vercel)"
direction TB
FE_Page["src/app/page.tsx"]:::frontend
FE_Layout["src/app/layout.tsx"]:::frontend
FE_Styles["src/app/globals.css"]:::frontend
subgraph "Server-side API Routes (hide x-api-key)"
direction TB
FE_API_Heroes["src/app/api/heroes/route.ts"]:::frontend
FE_API_Metadata["src/app/api/metadata/route.ts"]:::frontend
FE_API_Recommend["src/app/api/recommend/route.ts"]:::frontend
FE_Lib["src/lib/api.ts"]:::frontend
FE_Types["src/types/deadlock.ts"]:::frontend
end
end
%% AWS Inference Backend
subgraph "AWS Backend (HTTP API Gateway -> Lambda)"
direction TB
APIGW["API Gateway (HTTP API)\n/recommend /heroes /metadata"]:::aws
subgraph "Lambda (Python 3.11)"
direction TB
LH["backend/src/deadlock_backend/deploy/lambda_handler.py"]:::backend
IC["backend/src/deadlock_backend/inference/inference_core.py"]:::backend
CFG["backend/src/deadlock_backend/config.py"]:::backend
TMP["/tmp cache + module globals\n(warm reuse)"]:::backend
end
subgraph "Logs"
direction TB
CW["CloudWatch Logs + Alarm"]:::aws
end
subgraph "Private Model Assets"
direction TB
S3["S3 (private)\nmodel bundle + hero_assets + metadata"]:::storage
end
end
%% Offline training / export pipeline (runs outside request path)
subgraph "Offline Training + Bundle Export (local or CI runner)"
direction TB
NB["backend/notebooks/training_pipeline.ipynb"]:::pipeline
PIPE["backend/run_pipeline.py"]:::pipeline
subgraph "Export Scripts"
direction TB
EX_BUNDLE["backend/src/deadlock_backend/exports/export_inference_bundle.py"]:::pipeline
EX_HERO["backend/src/deadlock_backend/exports/export_hero_assets.py"]:::pipeline
EX_META["backend/src/deadlock_backend/exports/export_model_metadata.py"]:::pipeline
end
DW["Deadlock Parquet snapshots\n(S3 dump)"]:::external
DUCK["DuckDB\nfeature tables + joins"]:::pipeline
XGB["XGBoost training\n(early/mid/late/etc)"]:::pipeline
end
%% External Entities
User["User (Browser)"]:::external
GA["GA4 (analytics events)"]:::external
%% =========================
%% Runtime Request Flow
%% =========================
User -->|"HTTPS"| FE_Page
FE_Page -->|"fetch /api/* (server-side)"| FE_API_Recommend
FE_API_Recommend -->|"HTTPS + x-api-key\nPOST /recommend"| APIGW
FE_API_Heroes -->|"HTTPS + x-api-key\nGET /heroes"| APIGW
FE_API_Metadata -->|"HTTPS + x-api-key\nGET /metadata"| APIGW
APIGW -->|"Invoke"| LH
LH -->|"Cold start fetch"| S3
LH -->|"Warm reuse"| TMP
LH -->|"Logs/metrics"| CW
FE_Page -->|"analytics"| GA
%% =========================
%% Training + Export Flow
%% =========================
DW -->|"Read snapshots"| DUCK
NB -->|"Drive pipeline"| PIPE
PIPE -->|"Build features"| DUCK
DUCK -->|"Train"| XGB
XGB -->|"Export artifacts"| EX_BUNDLE
XGB -->|"Export metadata"| EX_META
EX_HERO -->|"Upload hero assets"| S3
EX_META -->|"Upload model metadata"| S3
EX_BUNDLE -->|"Upload inference bundle"| S3
%% =========================
%% Styles
%% =========================
classDef frontend fill:#B3E5FC,stroke:#0288D1,color:#000;
classDef backend fill:#C8E6C9,stroke:#388E3C,color:#000;
classDef aws fill:#E1BEE7,stroke:#6A1B9A,color:#000;
classDef storage fill:#FFE0B2,stroke:#F57C00,color:#000;
classDef pipeline fill:#D7CCC8,stroke:#5D4037,color:#000;
classDef external fill:#E0E0E0,stroke:#9E9E9E,color:#000;