AI-Powered Commmand Line Interface

Debug Infrastructure.
Not Dashboards.
Not Tickets.

Friday is the AI debugging CLI that runs entirely on your infrastructure. Multi-step diagnosis across gRPC, gNMI, YANG, and Linux kernel networking with full privacy by design.

View on GitHub Try DocLM Learn More
Friday — zsh
$ friday "why is pod payment-svc crashing in prod?"
▶ READFetching pod logs and events [payment-svc-7d9b8c]
▶ READChecking node resource pressure [node-03]
▶ ANALYZEOOMKilled — memory limit 256Mi, RSS 301Mi
▶ ANALYZECorrelating with 14:22 traffic spike (+340%)
▶ MODIFYRecommend: increase limits to 512Mi, add HPA
$

The problem

Your infra is screaming.
Your tools are silent.

Modern engineering teams drown in fragmented observability, alert fatigue, and knowledge gaps no dashboard solves.

01

Observability Fragmentation

Logs in Datadog. Traces in Jaeger. Metrics in Grafana. Correlating across tools costs 40% of incident response time.

02

Alert Fatigue

Teams receive 700+ alerts per day on average. Signal-to-noise ratio collapses. Real incidents get buried under noise.

03

Network Telemetry Complexity

gNMI streams, YANG models, and gRPC failures are notoriously hard to correlate. Diagnosing a silent drop requires deep protocol knowledge most engineers don't carry.

04

Data Privacy Blockers

Existing AI tools require shipping logs and configs to external APIs. Security teams block adoption on day one.

Execution model

Three-phase transaction engine.

Every debugging session follows a deterministic, reversible pipeline. No magic. No surprises.

User Journey
1
User
submits query
2
Gate 1
validation
3
RAG
retrieval
4
Prompt
construction
5
DocLM
inference
6
Gate 3
validation
7
Transaction
execution
Person
👤 User
friday "query"
one-shot mode

friday --it
interactive TUI
CLI Layer · Go + Cobra
GATE 1
Input Validation
Length bounds
UTF-8 validation
Injection pattern detect
Context Manager
Last 10 messages
Infra state retention
RAG Pipeline
MiniLM-L6-v2
ONNX · embedded in binary
Query embedding
Qdrant
Vector DB · Docker
Local similarity search
GATE 2
Quality Filter
Score ≥ 0.7
Top-5 chunks · diversity
Prompt Construction
Prompt Builder
System prompt
+ Function registry
+ RAG context chunks
+ Conversation history
+ Current user query
DocLM · vLLM (Docker)
Qwen2.5-Coder-3B
+ LoRA Adapter
Temperature = 0.1
Infra-fine-tuned
JSON Output
reasoning
execution_strategy
functions[]
Validation + Routing
GATE 3
Response Validation
JSON schema check
Function whitelist
Param type validation
Dep. graph analysis
Grounding score ≥ 0.6
↩ Self-Heal Retry
1x DocLM retry
on fixable errors
Variable Resolver
${prev.field}
${func[N].field}
Smart fallback
Transaction Engine
01 READ
Non-destructive queries
pod logs · net interfaces
DNS · metrics
02 ANALYZE
CPU-bound analysis
memory leak · YANG parse
core dump
03 MODIFY
Gate 4: dry-run
User confirm (before/after)
Snapshot → execute
Timeout + retry (max 2)
↩ LIFO Rollback on failure
DATA FLOW: LEFT → RIGHT  ·  ALL INFERENCE LOCAL  ·  NO DATA LEAVES YOUR INFRA
Gate 1 — Input Validation
Gate 2 — RAG Quality Filter
Gate 3 — Response Validation
Gate 4 — Dry-Run + User Confirm
LIFO Rollback
DocLM / LLM Boundary

Capabilities

Engineered for production reality.

Not a chatbot wrapper. A purpose-built debugging runtime with production-grade safety semantics.

Diagnostics

TCP & gRPC Health Checks

Runs ss -ti to extract live TCP state, retransmits, and queue depths. Connects to gRPC health endpoints and measures response latency end-to-end.

Streaming

gRPC Stream Analysis

Monitors streaming RPCs for a configurable window, tracks sequence gaps to detect packet drops, and surfaces flow control backpressure events with drop percentage.

Safety

Three-Phase Transaction Engine

Every session runs READ → ANALYZE → MODIFY in strict order. Destructive ops require dry-run validation and explicit confirmation before execution.

Kernel

Buffer Inspection & Sysctl Tuning

Reads /proc/sys/net/ buffer parameters, compares against bandwidth-delay product recommendations, and applies sysctl -w changes with automatic LIFO rollback on failure.

Chaining

Output Variable Resolution

Chain function outputs across steps using ${"{"}previous.port{"}"} syntax. The resolver navigates nested results and supports array indexing so each step builds on the last.

Privacy

Fully Local Inference

DocLM runs via vLLM inside your own infrastructure. No telemetry, no external API calls. Your network telemetry and YANG configs never leave your perimeter.

Fine-Tuned Model

Meet
DocLM.

A fine-tuned language model specialized in network telemetry diagnostics and debugging. Built on Qwen2.5-Coder-3B and merged into a single FP16 model — it is the inference engine behind Friday, purpose-built for network engineers working with gRPC, gNMI, YANG, and Linux kernel networking.

Qwen2.5
Base Model
On-Prem
Deployment
Network
Trained For
View on HuggingFace
DocLM v1 Ollama Runtime
  • Trained on network telemetry corpora: gNMI/gRPC specs, YANG models, Linux kernel networking docs, and incident post-mortems
  • Structured tool-use format: emits typed tool calls the agent executor resolves against live infra
  • Chains multi-hop reasoning: correlates symptoms across domains before proposing root cause
  • VRAM-efficient: runs on consumer GPUs and cloud spot instances alike
  • Open weights on HuggingFace for self-hosted enterprise deployment

Coverage

Every layer of your stack.

From L3 networking to container orchestration. Phase 1 ships networking deep. Phase 2 goes full-stack.

TCP Health CheckLive
DNSLive
Port ScanningLive
TracerouteLive
gRPC Health CheckLive
gRPC Stream AnalysisLive
Network Buffer InspectionLive
Sysctl TuningLive
Core Dump AnalysisLive
Variable ResolutionLive
State SnapshotsLive
Transaction EngineLive
gNMIcIncoming
K8s SupportIncoming
AWS SupportIncoming
GCP SupportIncoming

Privacy by design

Your logs stay in your infra. Always.

Friday deploys into your own cloud environment. No external API calls. No telemetry. Your incident data never leaves your perimeter.

Star on GitHub Try DocLM
Deployment Spec
LLM inference locationYour infrastructure
External API callsNone
Log data egressZero
Telemetry / trackingDisabled
Self-hosted LLMOllama / DocLM
Deployment modelCustomer cloud
Rollback on failureAutomatic LIFO