842 posts tagged with "ai-engineering"

Most AI teams treat a passing eval suite as a signal that their system is working. It isn't—not by itself. A suite that reliably scores 87% is doing exactly one thing: telling you the system performs well on the 87% of cases your suite happens to cover. If that suite was hand-curated six months ago, built from the examples the team thought of, and never updated against live traffic, it's measuring the wrong thing with increasing confidence.

This is the eval coverage problem. It's not about whether your evaluators are accurate—it's about whether the distribution of queries in your test set matches the distribution of queries your users are actually sending. When those two distributions diverge, you get a result that's far worse than a failing eval: a passing eval sitting on top of a silently degrading product.

The most common way teams schedule recurring AI agent jobs is also the most dangerous: a cron entry that fires a REST call every N minutes, which kicks off an LLM workflow, which either finishes or silently doesn't. This pattern feels fine in staging. In production, it creates a class of failures that are uniquely hard to detect, recover from, and reason about.

Cron was designed in 1975 for sysadmin scripts. The assumptions it encodes—short runtime, stateless execution, fire-and-forget outcomes—are wrong for LLM workloads in every dimension. Recurring AI agent jobs are long-running, stateful, expensive, and fail in ways that compound across retries. Using cron to schedule them is not just a reliability risk. It's a visibility risk. When things go wrong, you often won't know.

Your model was trained on data from last year. It was evaluated internally two months ago. It shipped a month after that. By the time a user hits a failure and you learn about it, you're already six months behind the world your model needs to operate in. This gap is not a deployment problem — it's a feedback loop problem. And most teams aren't measuring it, let alone closing it.

The instinct when a model underperforms is to blame the model architecture or the training data. But the deeper issue is usually the latency of your feedback system. How long does it take from the moment a user experiences a failure to the moment that failure influences your model? Most teams, if they're honest, have no idea. Industry analysis suggests that models left without targeted updates for six months or more see error rates climb 35% on new distributions. The cause isn't decay in the model — it's the world moving while the model stays still.

Your AI search feature launched three months ago. Early evals looked strong—your team ran 1,000 queries and saw 83% relevance. Thumbs-up rates were good. Users were engaging.

Then six weeks in, query reformulation rates started climbing. Session abandonment ticked up. A qualitative review confirmed it: users were asking different questions than they were before launch, and the model wasn't serving them as well as it used to.

Nothing changed in the model. Nothing changed in the underlying data. The product degraded because the users adapted to it.

This is the feedback loop trap. It is qualitatively different from the external concept drift most ML engineers train themselves to handle—and it is far harder to fix once it starts.

Most AI products ship with a thumbs-up/thumbs-down widget and call it feedback infrastructure. It isn't. What it is, in practice, is a survey that only dissatisfied or unusually conscientious users bother completing — and a survey that tells you nothing about what the correct output would have looked like.

The result is a dataset shaped not by what your users want, but by which users felt like clicking a button. That selection bias propagates into fine-tuning runs, reward models, and DPO pipelines, quietly steering your model toward the preferences of a tiny and unrepresentative minority. Implicit signals — edit rate, retry rate, session abandonment — cover every user who touches the product. They don't require a click. They're generated by the act of using the software.

Here's how to design feedback surfaces that produce high-fidelity training signal as a natural side effect of product use, and how to route those signals into your training pipeline.

Most teams figure out single-agent observability well enough. They add tracing, track token counts, hook up alerts on error rates. Then they scale to a hundred concurrent agents and discover their entire monitoring stack is watching the wrong things.

The problems that kill fleets are not the problems that kill individual agents. A single misbehaving agent triggering a recursive reasoning loop can burn through a month's API budget in under an hour. A model provider's silent quality degradation can make every agent in your fleet confidently wrong simultaneously — all while your infrastructure dashboard shows green. These failures don't show up in latency charts or HTTP error rates, because they aren't infrastructure failures. They're semantic ones.

Most teams reach for vector embeddings when building RAG pipelines. It's the obvious default: embed documents, embed queries, find the nearest neighbors, feed results to the LLM. It works well enough on the demos. Then they deploy to a compliance team or a scientific literature corpus, and accuracy falls off a cliff. Not gradually — abruptly. On queries involving five or more entities, vector RAG accuracy in enterprise analytics benchmarks drops to zero. Not 50%. Not 20%. Zero.

This isn't a configuration problem. It's an architectural mismatch. Vector retrieval treats documents as points in semantic space. Knowledge graphs treat them as nodes in a relational structure. When your queries require traversing relationships — not just finding similar content — the topology of your retrieval architecture is what determines whether you get the right answer.

Most teams add human-in-the-loop review as an afterthought: the agent finishes its chain of work, the result lands in a review queue, and a human clicks approve or reject. This feels like safety. It is mostly theater.

By the time a multi-step agent reaches end-of-chain review, it has already sent the API requests, mutated the database rows, drafted the customer email, and scheduled the follow-up. The "review" is approving a done deal. Declining it means explaining to the agent — and often to the user — why nothing that happened for the past 10 minutes will stick.

The damage from misplaced approval gates isn't always dramatic. Often it's subtler: reviewers who approve everything because the real decisions have already been made, engineers who add more checkpoints after incidents and watch trust in the product crater, and organizations that oscillate between "too much friction" and "not enough oversight" without ever solving the underlying placement problem.

A common assumption in AI product development goes something like this: if a model can handle a hard task, it can definitely handle an easier one nearby. This assumption is wrong, and it's responsible for a category of production failures that no amount of benchmark reading prepares you for.

The research term for the underlying phenomenon is the "jagged frontier" — AI's capability boundary isn't a smooth line that hard tasks sit outside of and easy tasks sit inside. It's a ragged, unpredictable shape. AI systems can write production-grade database query optimizers and still miscalculate whether two line segments on a diagram intersect. They can pass PhD-level science exams and fail children's riddle questions that involve spatial relationships. They can synthesize 50-page documents and then confidently hallucinate a summary of a paragraph they just read.

Product teams have started routing analytical questions directly to LLMs: "What's causing the churn spike?" "Why did conversion drop after the redesign?" "Which cohort should we focus retention spend on?" The outputs land in executive decks, drive roadmap decisions, and get presented to investors. The models answer confidently, in polished prose, with specific numbers. And a significant fraction of those answers are wrong in ways that don't announce themselves.

This isn't a general criticism of LLMs for data work. There are tasks where they genuinely help. The problem is that the failure modes are invisible — the model doesn't hedge, doesn't caveat, and doesn't distinguish between "I computed this from your data" and "I generated something that sounds like what this number should be." Practitioners who understand where the breakdowns happen can capture the genuine value and route around the landmines.

Model routing is the first optimization most teams reach for. Route simple queries to a small cheap model, complex ones to a large capable model. It works well for managing cost and throughput. What it cannot fix is the wall you hit when the physics of cloud inference collide with a latency requirement of 100ms or less. A network round-trip from a mid-tier data center already consumes 30–80ms before a single token is generated. At that point, routing is irrelevant — you need to either run the model closer to the user or run a substantially smaller model. Both paths require compression decisions that most teams approach without a framework.

This is a guide for making those decisions. The three techniques — quantization, knowledge distillation, and on-device deployment — solve overlapping problems but have very different cost structures, quality profiles, and operational consequences.

Your per-request error rates look clean. Latency is within SLO. The LLM judge is scoring outputs at 87%. And then a user files a support ticket: "I told the bot my account number three times. It just asked me again." A different user: "It agreed to a refund, then two turns later denied the policy existed."

Single-turn failures are visible. The request comes in, the model hallucinates or refuses, your eval catches it, you fix the prompt. The feedback loop is tight. Multi-turn failures work differently: the session starts fine, degrades gradually turn by turn, and your monitoring never fires because each individual response is technically coherent. The problem is the session as a whole — and almost no team instruments for that.

Research across major frontier models (Claude 3.7 Sonnet, GPT-4.1, Gemini 2.5 Pro) shows an average 39% performance drop when moving from single-turn to multi-turn conversations. That number hides the real story: only about 16% of the drop is capability loss. The other 23 points are a reliability crisis — the gap between a model's best and worst performance on the same task doubles as conversation length grows. You're not just getting worse outputs; you're getting inconsistent ones.