Loading…
Topic
18 posts about Human-in-the-Loop. Every summary links to the original.
See the idea
AI-DLC is an AI-driven development lifecycle that runs delivery through Inception, Design, and Construction, with specs, steering files, and hooks keeping agents aligned to agreed intent.
AI safety is a budget: prevention, verification, and containment each buy different protection at a different cost. This map shows what each control costs and when it pays off.
Asset management runs as a layered control plane: custodian and pricing data feed a governed context layer, agents run the overnight fund work, and a human signs off before anything is published.
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
Newest first
A production LLM system must know when it is uncertain, not just produce answers. The confidence layer composes a score from four independent signals — model self-score, deterministic verification, judge agreement, and historical accuracy — with the model’s own claim deliberately weighted lowest. That raw score is calibrated via isotonic regression against empirical accuracy so that a score of 0.9 means right roughly 90% of the time. A per-slice threshold decides automation versus human review, and an abstain floor escalates inputs too uncertain to even draft. An independent judge evaluates high-stakes decisions, feeding its agreement into the confidence score and flagging disagreements. Human handoffs show the draft, reasoning, evidence, and why it was routed, and every human action is captured as signal to tune thresholds and improve the agent.
Human-in-the-loop (HITL) frameworks integrate human judgment into automated AI workflows to improve reliability and safety, but without careful design they can create bottlenecks. The post reveals key trade-offs: more automation increases speed and reduces cost but raises risk, while more human oversight boosts accuracy and safety but adds latency and cost. Common failure modes include over-reviewing trivial cases, lacking clear trigger logic, and expecting human reviewers to scale linearly. The proposed solution is a tiered HITL system applying the Pareto principle: automation handles 85% of high-confidence decisions using confidence calibration, with three tiers addressing the remaining cases under 3 ms latency, asynchronous review within hours, and real-time oversight with 30-60 second windows. Success requires careful routing, reviewer tooling, and closing feedback loops across seconds, days, weeks, and quarterly cycles to continuously improve model performance and reduce human involvement over time.
Claude Managed Agents runs Anthropic’s Claude as a hosted agent without managing inference infrastructure. Pairing it with MongoDB Atlas provides fast access to operational and vector data through a single pymongo connection, eliminating the need for separate vector, full-text, and graph stores. This tutorial builds a human-in-the-loop fraud review agent that retrieves flagged transactions, runs four retrieval patterns (vector, full-text, hybrid, and graph) against one collection, and escalates risky cases to a human reviewer using Managed Agents’ native pause. The credential-safe integration uses host-side custom tools so the MongoDB connection string never enters the agent sandbox, keeping secrets on your side. The agent verifies a signed authorization mandate, uses $graphLookup to detect mule chains and $rankFusion for hybrid search, writes immutable verdicts and audit events back to the same cluster, and triggers durable human pauses on structuring deposits just under $5,000, wires over $50,000, and ring signals from graph traversal.
Organizations often retain specialist knowledge in experts’ heads, leaving routine compliance assessments slow, inconsistent, and difficult to scale. The proposed “second brain” combines a structured, auditable knowledge architecture with a separate reasoning layer and a self-improvement loop that converts expert feedback into verified, regression-tested text updates without model retraining. High-density organizational positions, taxonomies, routing indexes, and gateway files are maintained in a dependency-linked filesystem, while sparse supporting material remains available through semantic or lexical retrieval. Composable recipes encode expert procedures, and an evaluation pipeline proposes auditable pull requests, retries failed fixes, and expands regression coverage after approval. After six weeks, SMEs reportedly found outputs useful almost always, assessments fell from days to minutes, and improvement cycles produced zero regressions; the architecture is presented as applicable beyond compliance.
The FDA built HALO (Harmonized AI and Lifecycle Operations for Data) as a secure, governed, AI-ready enterprise data platform for modernizing siloed systems without interrupting regulatory work. Its move to Databricks on AWS GovCloud, following FedRAMP High authorization sponsorship, added Unity Catalog as a governance layer across a multi-tenant architecture, while Terraform-based security patterns, PrivateLink, customer-managed keys, and the compliance security profile support regulated workloads. The agency migrated more than 5,000 users and 8,000 jobs and pipelines with zero downtime, refactoring over 1,000 pipelines and 4,000 notebooks. After onboarding eight centers and 30 programs, FDA reported query responses improving over 30%, compute costs falling over 20%, and provisioning and sharing time dropping over 75%. HALO also supports responsible AI use cases such as MARS, with humans retaining decision authority.
Twilio research cited in the post finds that 78% of consumers have tried to bypass an AI agent to reach a human, while 63% want easy escalation and only 47% of brands provide it. The reported causes are unclear AI disclosure, discomfort with high-stakes tasks, and prior failures: 90% of tested consumers misidentify text-based AI, 80% report a negative AI experience, and failed attempts add 81 seconds before escalation. Consumers are most comfortable with scheduling and reservations, returns and exchanges, and concert seat selection, but pull back on medical inquiries, payments, and loan applications. The proposed operating model combines routine automation with clear disclosure, obvious escalation, persistent customer memory, and real-time detection of sentiment and escalation risk. It argues that trust, rather than wholesale replacement, should determine how humans and AI divide customer-service work.
The article outlines authentication trends expected to shape identity verification, fraud prevention, payments, and access in 2026, against a backdrop of deepfakes, synthetic identities, regulatory change, and increasingly online services. It describes identity orchestration that combines document checks, biometrics, device intelligence, and behavioral signals, while fraud systems assess behavior and may allow, challenge, delay, or block access. For AI agents, Know Your Agent links a bot to its enterprise and official communication channel, with document verification and Twilio Lookup presented as a combined onboarding approach. The recommended strategies add continuous anomaly monitoring, risk-based MFA, human-in-the-loop approvals, multichannel authorization, passkeys, digital IDs, and layered controls, while concluding that businesses and regulators should remain agile and use services such as Twilio Verify, Lookup API, and Verify Fraud Guard.
Dynamic catalogs in food delivery and e-commerce make knowledge-graph (KG) maintenance difficult: LLM- or scraper-generated relationships can be semantically wrong, manual review does not scale, and bad edges can degrade query expansion, CTR, and trust. The proposed feedback-driven verification engine treats unverified edges as hypotheses and uses the search interface as a validation environment. It separates verified and candidate edges, injects a limited number of candidates into lower-risk interface positions through an exploration-versus-exploitation policy, and records context-anchored user interactions with weighted positive and negative signals. An offline engine aggregates exposure and engagement into a confidence ratio, promoting candidates above a verification threshold and pruning those below a failure threshold. The framework is presented as a continuous human-in-the-loop method for validating parent-child and sibling relationships at scale; planned extensions add confidence tiers and defenses against spam, bots, and Sybil attacks.
Human-in-the-loop (HITL) is an AI and machine learning approach that places people in training, supervision, or decision-making to improve accuracy, safety, and ethical alignment. Its feedback loop can include data labeling, output review, escalation, approval, override, and continuous feedback, with confidence thresholds and risk scoring routing only selected decisions to people. The explainer distinguishes HITL, where review occurs before flagged actions, from human-on-the-loop monitoring and human-over-the-loop governance, and separates HITL from RLHF, a training-specific technique. It describes uses in medical imaging, moderation, autonomous vehicles, financial services, and AI agents handling consequential actions. Databricks Agent Bricks is presented as supporting governed traces and Agent Learning from Human Feedback, including a case where 32 feedback items improved instruction-following from roughly 12% to 80%.
CPS asset identification is difficult because 88% of CPS assets do not transmit an exact product code and 76% use product codes that differ from the vendor's official records. Claroty's AI-Powered CPS Library addresses this entity-resolution problem with a hybrid architecture that combines classic matching, statistical inference, generative AI, NLP, reasoning agents, and human-in-the-loop review to reconcile protocol data, vendor records, manuals, and vulnerability information. Its Databricks-based lakehouse uses Delta Lake, Unity Catalog, Spark pipelines, Model Serving, MLflow evaluations, Lakebase, and Databricks Apps to maintain governed, auditable mappings across a catalog of more than 17 million assets. The post reports 25% improved vulnerability-attribution accuracy and says early tests gave 56% of analyzed devices new or updated recommendations for outdated firmware, while low-confidence mappings remain subject to expert review.
Uber built an AI-powered PRD Evaluator to help product managers strengthen drafts before they enter structured checkpoint and approval forums. Starting with a PRD, it gathers linked documents, decks, meeting notes, prior experiments, cross-functional artifacts, and Uber-specific context, then classifies the proposal to calibrate review depth. It assesses opportunity and hypothesis, product scope, user experience and impact, and metric and data rigor, producing a launch-readiness scorecard with prioritized gaps, evidence, replacement-text suggestions, and a “start here” fix. Early internal use by dozens of PMs helped surface blind spots, test unsupported headroom assumptions, identify adjacent-system effects, and improve experiences within existing scope. The team’s lessons emphasize decision frameworks, broader context, hard boundaries, prioritization, and preserving human judgment for final approval.
Public-sector agencies face fraud methods including synthetic identities, deepfake-enhanced documentation, and personalized social engineering, while legacy controls remain fragmented and difficult to scale. The fictional Services Bureau demonstrates an operating model that combines Databricks Apps, Lakebase, Unity Catalog, Delta Sharing, Agent Bricks, and AI/BI Genie in a single fraud-operations environment. Governed data lands in Delta tables; Unity Catalog applies attribute-based access control, masks PII by role, and provides lineage, while agents connect live lakehouse queries, agency policies, and external fraud signals through MCP. Analysts review evidence and recommendations, then approve, override, or escalate cases, keeping human judgment central. The described workflow turns weeks of manual investigation into a day, supports dashboards and conversational SQL, and is presented as making fraud decisions faster, more secure, transparent, and defensible.
Uber describes its deactivation principles and safeguards for drivers and couriers, motivated by the stress of losing app access and the need for decisions that are fair, accurate, and transparent. Most access issues are temporary, while permanent deactivations are described as rare and intended to protect platform safety; users should receive notice, relevant details, and an opportunity to respond whenever possible. The updated process lets drivers view eligible reports, submit context including Record My Ride audio or video, and have reports reviewed alongside their information, while systems filter suspected fraudulent or false claims. The Review Center explains deactivation reasons and appeal options, supports uploads, and sends every appeal to a real person, with additional measures including excluding low ratings caused by factors outside a driver's control, limiting trip categories when appropriate, and covering third-party testing for disputed impairment claims.
The post frames AI as a new “miracle material” that could reshape knowledge work as steel and semiconductors reshaped earlier eras. It argues that coding agents already let one programmer orchestrate several AI coding agents, but broader knowledge work remains constrained by fragmented context across tools and weak verifiability. At the organisational level, it compares AI with steel and steam, suggesting agents could preserve workflow context, surface decisions, and reduce dependence on meetings and hierarchical approvals; Notion is described as using more than 700 agents alongside 1,000 employees for repetitive tasks. At economy scale, it predicts organisations spanning humans and thousands of agents, with continuous workflows and new rhythms, while acknowledging that this transition may be disorienting and requires redesign rather than simply adding chatbots to existing tools.
During a fall internship at Ramp, the author worked on reducing repetitive accounting work around month-end close, focusing on amortization and ERP mapping workflows. One project uses receipt data, merchant category codes, customer history, and embedding similarity to detect likely prepaid expenses and suggest service periods and accounts or templates, with recommendations reviewable before application. A second built Ramp-Native Amortization to create schedules, generate periodic debit and credit journal entries, handle catch-ups and reclassifications, and sync entries to ERPs through a dashboard; transaction and reimbursement support was nearing alpha production release, while bill pay was next. The third generates high-confidence mapping-rule suggestions from customer coding patterns on a nightly schedule. Across the work, backtesting, accuracy metrics, and human workflow research support automation intended to earn trust rather than remove accountant judgment.
Debug Mode is an agent loop for coding bugs that combines runtime information with human verification, addressing issues that can stump conventional coding agents. After the user describes a bug, it reads the codebase, proposes multiple hypotheses, and instruments the code with logging statements designed to test them. The user reproduces the issue while the agent collects variable states, execution paths, and timing information, then uses those logs to identify a root cause and produce a targeted fix, often a two- or three-line modification. The user reproduces the bug again with the fix; if it works, the agent removes its instrumentation, while a failed verification prompts additional logging and refinement until the user considers the problem solved.
Ramp describes how it built and shipped LLM-backed agents for expense management, focusing on the trust risks created by low-quality or unexpected finance outputs. It recommends choosing ambiguous, high-volume problems with asymmetric upside, then combining reasoning and citations grounded in product or user-provided context with an explicit “Needs review” path when the agent is unsure. The policy agent uses expense policies to approve, reject, or escalate expenses, while deterministic rules such as dollar limits, vendor blocklists, and category restrictions constrain autonomy. Ramp reports that more than 65% of approvals are fully handled by the agent, while users can progressively move from suggestions to autonomous actions through workflow-based controls and evaluation.
Ramp describes principles for building customer-focused AI, contrasting embedded outcomes with superficial chatbot experiences it calls “AI washing.” The approach emphasizes integrating models into workflows, separating general models trained on aggregated and masked data from sensitive models that use private customer data temporarily through in-context learning, and requiring explicit opt-in for sharing contract data. It prioritizes user control and feedback over lengthy explanations, while using guardrails that constrain outputs through predefined interactive data blocks rather than relying only on censorship filters. Ramp cites contract term extraction, spend intelligence, pricing benchmarks, Copilot, and Jsonformer as examples of these practices. The stated goal is AI that saves customers time and money while protecting privacy and producing safer, more useful results.
The best engineering reads, once a week.
