FRAMEWORM

Production-Grade Generative AI Framework

The Product Family

FRAMEWORM is not a single tool. It is four products that cover the full ML lifecycle — from training to production.

FRAMEWORM          → core ML infrastructure
FRAMEWORM AGENT    → autonomous training monitor
FRAMEWORM SHIFT    → distribution drift detection
FRAMEWORM DEPLOY   → one-command model deployment

FRAMEWORM (Core)

Production-grade ML infrastructure. Dependency graph engine, Bayesian HPO, mixed precision training, TorchScript + ONNX export, and 6 built-in generative architectures.

Built-in architectures: VAE · DCGAN · DDPM · VQ-VAE-2 · ViT-GAN · CFG-DDPM

pip install frameworm
frameworm init my-project
cd my-project
frameworm train --config config.yaml

FRAMEWORM AGENT

Autonomous training monitor. Watches loss curves and gradient norms in real time, detects anomalies, and intervenes automatically.

frameworm agent start --experiment my_run

Benchmark — 10 Experiments

Each run paired with a shadow run (no agent) as counterfactual comparison.

#   Experiment           Condition        Agent Loss   Shadow Loss    D Loss   Resolved
─────────────────────────────────────────────────────────────────────────────────────────
1   Baseline-1           baseline            70.39        70.37        -0.02       -
2   Baseline-2           baseline            72.61        70.35        -2.27       -
3   Baseline-3           baseline            72.48        70.39        -2.09       -
4   HighLR-1             lr=8e-02            75.71      9999.00     +9923.29      YES
5   HighLR-2             lr=5e-02            69.99        70.37        +0.37      YES
6   GradExplosion-1      grad_explosion      75.93        70.35        -5.58      YES
7   GradExplosion-2      grad_explosion      76.07        70.35        -5.72      YES
8   LowLR-Plateau-1      lr=1e-06            94.65       283.47      +188.82      NO*
9   LowLR-Plateau-2      lr=5e-07            95.31       285.57      +190.26      NO*
10  AgentIntervenes      lr=1e-01            75.84      9999.00     +9923.16      YES
─────────────────────────────────────────────────────────────────────────────────────────
Baseline avg loss:              71.83
Runs with anomalies detected:   7 / 10
Anomalies resolved by agent:    5 / 7
Resolution rate:                71.4%
False positives on clean runs:  0 / 3

Key results:

• Experiments 4 and 10: agent loss ~75 vs shadow loss 9999 — runs that would have fully diverged were saved
• Experiments 6 and 7: gradient explosion detected and recovered, final loss within 8% of healthy baseline
• 0 false positives on clean baseline runs

*Plateau experiments (8, 9): agent detected and intervened 16 times each, reducing loss from ~285 to ~95 (+190 improvement). Full convergence within epoch budget is a known limitation and direction for future work.

FRAMEWORM SHIFT

Drop-in distribution drift detection. Two lines to attach to any model in production.

from frameworm.shift import ShiftMonitor

monitor = ShiftMonitor.from_reference("experiments/my_model.shift")
monitor.check(incoming_batch)   # fires alert automatically if drift detected

Three integration surfaces:

# SDK
monitor = ShiftMonitor("my_model")
monitor.check(batch)

# FastAPI middleware — zero changes to your endpoint
app.add_middleware(ShiftMiddleware, reference="my_model")

# CLI
frameworm shift check --reference train.csv --current live.csv
frameworm shift report --reference train.csv --current live.csv --output report.html

Severity levels: NONE · LOW · MEDIUM · HIGH Alerts via: Slack · webhook · log file · stdout

FRAMEWORM DEPLOY

One command from trained model to production API.

frameworm deploy start --model experiments/checkpoints/best.pt --name my_model

What that single command does:

• Exports model to TorchScript + ONNX
• Generates a model-aware FastAPI server (correct schema per architecture)
• Builds a multi-stage Docker image with HEALTHCHECK baked in
• Starts p50/p95/p99 latency tracking on every request
• Auto-attaches FRAMEWORM SHIFT drift monitoring
• Starts a background rollback controller

Auto-rollback: p95 latency spike for 3 consecutive checks → stops bad container, starts previous known-good version, fires Slack alert. No human needed.

Every deployed model exposes:

GET /predict   → inference
GET /health    → liveness
GET /ready     → readiness
GET /metrics   → p50/p95/p99 + error rate

frameworm deploy stop     --name my_model
frameworm deploy status   --name my_model
frameworm deploy rollback --name my_model
frameworm deploy promote  --name my_model --version v2.0 --stage production

Model-aware server generation. FRAMEWORM DEPLOY knows all 6 built-in architectures and generates architecture-specific inference code. A DDPM server runs a 1000-step denoising loop. A VAE server unpacks three return values. A CFG-DDPM server accepts guidance scale. Generic tools can't do this.

Why FRAMEWORM?

Quick Start

pip install frameworm

# Train a model
frameworm train --config config.yaml

# Monitor training autonomously
frameworm agent start --experiment my_run

# Check for drift in production
frameworm shift check --reference train.csv --current live.csv

# Deploy to production
frameworm deploy start --model experiments/checkpoints/best.pt --name my_model

Documentation

• Quick Start
• FRAMEWORM AGENT
• FRAMEWORM SHIFT
• FRAMEWORM DEPLOY
• API Reference

License

MIT — see LICENSE

Citation

@software{frameworm2026,
  title  = {FRAMEWORM: Production-Grade Generative AI Framework},
  author = {Aakash Ali},
  year   = {2026},
  url    = {https://github.com/Aakash0440/frameworm}
}