Models
Introducing afaxxaq-000 — unname
A pre-release note on afaxxaq-000 — unname, the open-weight multimodal model Navin Research is testing and preparing to announce.
Notes
Latest notes
Latest notes
Featured note
Models
Moyi: Speech Technology With the Boundary Visible
How Moyi connects acoustic context, speech transformation, local runtime, and human review without hiding the unknowns.
Agents
Start With One AI Agent Before Building a Multi-Agent System
A practical decision rule for choosing one agent, a workflow, or several collaborating agents.
Older notes
Older notes
Agents
Multi-Agent Orchestration Patterns That Work in Production
Choose among manager-worker, parallel fan-out, handoff, and evaluator loops without building an uncontrolled agent swarm.
Agents
How to Design Reliable Handoffs Between AI Agents
Use typed handoff packets, acceptance checks, and clear responsibility boundaries between specialist agents.
Agents
Design Better Tools for AI Agents
Make agent tools easier to select, call, validate, and recover from with narrow schemas and actionable errors.
Agents
AI Agent Memory: What to Store and What to Recompute
Separate working context, durable user facts, episodic history, and external knowledge to reduce cost and privacy risk.
Agents
Secure an MCP Server Before Connecting AI Agents
A deployment checklist for authentication, authorization, consent, validation, and least-privilege Model Context Protocol tools.
Agents
A2A vs MCP: Different Layers for Agent Systems
Understand when Agent2Agent and Model Context Protocol complement each other instead of competing.
Agents
Human Approval Gates for High-Risk AI Agent Actions
Place approval at the last responsible moment and show people exactly what the agent intends to do.
Agents
Build Long-Running AI Agent Jobs That Can Resume Safely
Use checkpoints, idempotent tools, bounded retries, and cancellation for tasks that outlive one request.
Agents
AI Agent Observability: Traces That Help Debug Failures
Capture decisions, tool calls, latency, cost, and evidence without turning traces into a privacy liability.
Reasoning
Reasoning Models vs Chat Models: A Practical Selection Guide
Choose the least expensive model behavior that reliably solves the task instead of enabling heavy reasoning everywhere.
Reasoning
Route Reasoning Effort by Task Difficulty
Build a simple difficulty router with escalation based on confidence and verification rather than one expensive default.
Reasoning
Verify AI Reasoning Without Requesting Chain-of-Thought
Check final answers with tests, constraints, citations, and independent computation instead of collecting private reasoning traces.
Reasoning
Structured Outputs for Reliable AI Pipelines
Use JSON Schema for machine-consumed responses while preserving server-side validation and business rules.
Reasoning
Reasoning With Tools: Separate Planning From Execution
Let models choose actions while deterministic code validates, executes, and records each operation.
Reasoning
Long Context Is Not the Same as Memory
Use large context windows deliberately; retrieval, summaries, and durable state still matter.
Reasoning
A Context Engineering Checklist for AI Agents
Curate instructions, tools, evidence, state, and history as a limited resource on every model turn.
Reasoning
Build a Small Reasoning Evaluation Dataset That Matters
Create a compact, versioned eval set from real failures before chasing broad benchmark scores.
Models
How to Choose an AI Model for a Production Application
Evaluate quality, latency, cost, context, tool use, deployment, and governance with your own workload.
Models
Open-Weight vs Hosted AI Models: A Deployment Decision
Compare control, operations, licensing, data handling, and performance before self-hosting or buying an API.
Models
Integrate the DeepSeek API Without Locking Your Application
Use an adapter, capability checks, strict validation, and live documentation for a resilient DeepSeek integration.
Models
A Deployment Checklist for Kimi Open-Weight Models
Verify the model card, hardware fit, serving stack, tool format, and license before deploying Kimi models.
Models
A Practical Qwen Model Deployment Checklist
Select the right Qwen variant, template, precision, and serving engine while preserving reproducibility.
Models
Self-Hosting Llama: What to Validate Before Production
Cover licensing, weights, serving, safety, evaluation, and operations for a responsible Llama deployment.
Models
Reliable Function Calling With Mistral Models
Build a provider adapter, validate arguments, and test tool selection with official Mistral capabilities.
Models
An AI Model Upgrade Playbook With Safe Rollback
Change models through pinned versions, regression evals, shadow traffic, canaries, and reversible configuration.
Multimodal
Build a Reliable Image Understanding Pipeline
Prepare images, preserve provenance, structure outputs, and verify visual claims before using multimodal results.
Multimodal
Extract Structured Data From Documents With Multimodal Models
Combine file parsing, page images, schemas, citations, and validation for dependable document extraction.
Multimodal
Practical Video Understanding With Multimodal AI
Sample video deliberately, keep timestamps, separate audio and visual evidence, and evaluate event recall.
Multimodal
Audio Understanding Beyond Transcription
Design an audio pipeline for transcription, speaker turns, acoustic events, and grounded summaries.
Multimodal
Build Safer Real-Time Voice Agents
Handle latency, interruptions, turn detection, tool calls, consent, and fallback in live voice systems.
Multimodal
Multimodal RAG for Text, Images, and Tables
Index each modality with provenance, retrieve at the right granularity, and cite original evidence.
Multimodal
Defend Multimodal Agents Against Visual Prompt Injection
Treat text inside images and documents as untrusted data, not system instructions.
Multimodal
Evaluate Multimodal AI With Evidence-Level Metrics
Measure perception, grounding, extraction, reasoning, and abstention separately.
Inference
Reduce LLM Latency Before Changing Models
Cut sequential calls, shorten outputs, stream useful progress, cache stable prefixes, and parallelize independent work.
Inference
Prompt Caching: Structure Requests for Reuse
Place stable prefixes first, separate volatile data, monitor cache hits, and protect tenant boundaries.
Inference
When to Use Batch Inference for AI Workloads
Move delay-tolerant, independent jobs to batches while preserving identifiers, retries, and quality checks.
Inference
Quantization for LLM Serving: Measure the Quality Trade-Off
Reduce memory and increase throughput only after evaluating the exact quantized artifact on your tasks.
Inference
KV Cache Planning for LLM Inference
Estimate memory from context, concurrency, model architecture, precision, and reuse before promising capacity.
Inference
Prefill and Decode Disaggregation: When It Helps
Separate compute-heavy prompt processing from memory-bandwidth-heavy token generation only for the right workload and scale.
Inference
Cost-Aware Model Routing Without Silent Quality Loss
Route tasks among models using validated task classes, hard constraints, and escalation rather than price alone.
Inference
Benchmark an LLM Inference Server Correctly
Report workload, hardware, concurrency, prompt and output lengths, latency percentiles, throughput, and quality together.
Safety Evaluation
Prompt Injection Defense for Tool-Using AI Agents
Combine trust separation, least privilege, validation, approvals, and monitoring because prompts alone cannot secure agents.
Safety Evaluation
Evaluate AI Agents End to End
Measure task success, policy compliance, tool behavior, efficiency, and recovery across complete trajectories.
Safety Evaluation
Build Tool-Call Evaluations for AI Agents
Test tool selection, arguments, sequencing, authorization, recovery, and unnecessary calls with executable fixtures.
Safety Evaluation
Evaluate RAG Before Tuning the Generator
Separate retrieval recall, ranking, context quality, answer support, and citation correctness.
Safety Evaluation
Use LLM-as-a-Judge Carefully
Design rubrics, randomize ordering, calibrate against humans, and keep deterministic checks first.
Safety Evaluation
Red-Team an AI Agent With Realistic Attack Paths
Test untrusted content, tool abuse, data boundaries, identity, and recovery in a controlled environment.
Safety Evaluation
Privacy Engineering for AI Applications
Minimize data, define retention, isolate tenants, control logs, and make deletion testable.
Safety Evaluation
Continuous AI Evaluations in CI and Production
Run fast regression gates on every change and monitor sampled real-world outcomes without leaking user data.