Route Simple Agent Steps Cheaply—and Escalate with Evidence
Compare pre-generation model routing with confidence-driven cascades using quality, cost, tail latency and end-to-end agent task success—not provider failure fallback.
Algotechnic journal
Production ML,
examined in practice.
Engineering notes on designing, deploying and operating dependable machine learning systems, with a focus on practical constraints and operational value.
2026
Agent Memory Needs a Lifecycle, Not Just Retrieval
Agent memory needs provenance, temporal validity and deletion as well as retrieval. Separate evidence from facts, correction from erasure and recency from expiry.
A Least-Privilege Boundary for Agent Tools
Design agent tool security around expiring grants, typed mediation, isolated execution, per-action authorisation, selective approval and useful audit evidence.
Evaluate the Trajectory, Not Just the Final Answer
Build agent evaluations around action logs, state transitions, subgoal progress, recovery and side-effect checks without enforcing one golden path.
Agentic Workflows That Know When to Stop
A single-agent run needs explicit completion checks, bounded retries, error exits and useful human handoffs; autonomy alone does not make tool use reliable.
2025
AI Evals and Traces: Turn Agent Failures into Regression Tests
Use agent traces and domain feedback to build labelled failure cases, layered offline evals, release gates and production monitoring without mistaking logs for truth.
Context Engineering: Select, Compress, and Verify What the Agent Sees
Manage an agent’s token and state lifecycle with selective retrieval, lossy compaction, artefact isolation and tests for omitted or contradictory evidence.
Repository-Scale Coding Assistants Need Specifications, Tests, and Review
Codex and Copilot can turn issues into pull requests, but useful delegation depends on precise specifications, repository context, test evidence and human review.
Model Context Protocol: Capability Contracts and Security Boundaries
Model Context Protocol standardises AI tool and context exchange, but capability negotiation cannot replace consent, validation, isolation or authorisation.
Reasoning Models and Inference Compute: What o1 and DeepSeek R1 Made Visible
How RL teaches longer reasoning trajectories, why verification remains domain-bound, and how to balance accuracy against latency and cost per solved task.
2024
Mixture-of-Experts in Production: Sparse Compute, Dense Operational Constraints
How top-k routing, expert capacity, communication and total-weight memory determine whether sparse MoE models deliver useful production throughput.
Structured Outputs and Interfaces: Enforce Schemas Without Surrendering Validation
Schema-constrained generation makes model output machine-readable, but production systems still need semantic checks, bounded retries and authorised tool execution.
Small Language Models in Production: Fit the Model to the Task and Hardware
Choose small language models through domain tests, quantised quality deltas and hardware measurements, with adapters, privacy controls and fallback designed explicitly.
Long-Context Retrieval: Selection Still Matters When Everything Fits
Long context windows do not remove the need for retrieval. Compare answer quality, position robustness, latency, freshness and access control.
Multimodal Models in Production: Evaluate Text, Images and Documents Together
Build multimodal evaluation around faithful preprocessing, document layout, cross-modal reasoning, grounded answers and the cost of each successful task.
2023
LLM Evaluation in Production: Task Sets, Human Review, and Regression Gates
Build a versioned LLM evaluation suite from real tasks and failures, then gate releases with deterministic tests, calibrated judging, human review, cost and latency.
LoRA and QLoRA for Open-Weight Models: Prompting First, Fine-Tuning with Evidence
How low-rank adapters and a frozen quantised base reduce adaptation costs, and when repeatable behaviour justifies fine-tuning instead of prompting.
Vector Search as Application Infrastructure: Embeddings, ANN Indexes, Filters, and Operations
Embeddings become application infrastructure only when index design, metadata filters, retrieval quality and operational ownership are engineered as one system.
Retrieval-Augmented Generation: Grounding, Freshness, and the Limits of Citation
Follow retrieval-augmented generation from ingestion and embeddings to top-k evidence and answer synthesis, with separate checks for freshness, support and attribution.
GPT-4 Model Evaluation: Capabilities, Uneven Performance, and Benchmark Limits
GPT-4’s benchmarks show major capability gains, but contamination checks, prompting and post-training complicate claims of reliable, uniform application performance.
2022
ChatGPT: Dialogue, RLHF and the Limits of a Public Research Preview
How ChatGPT used dialogue demonstrations and ranked responses to make a language model accessible, and why fluent conversation still demands verification.
Instruction Tuning and RLHF: Teaching Language Models to Follow Instructions
FLAN and InstructGPT take different routes to instruction following. Compare supervised task generalisation with human-feedback alignment, evaluation and regressions.
Diffusion Models: Denoising, Guidance and the Text-to-Image Breakthrough
How iterative denoising, guidance and two distinct conditioning designs shaped unCLIP and Imagen, and why sample quality brings a substantial inference bill.
Why Transformers Became the Default Architecture for Foundation Models
How attention, parallel training and pretrain–transfer made Transformers the leading foundation-model architecture, without eliminating CNNs or inference bottlenecks.
MLOps Is an Operating Model for the Whole ML Lifecycle
MLOps connects telemetry, reproducible releases and clear ownership across the ML lifecycle, turning drift alerts into validated operational decisions.