Recommendation · MoneyHarish Gaggar recommends implementing agent PII treatment because queries can accidentally ingest secrets — e.g., uploading a doc with sensitive information or typing API keys. Put protections in place so secrets do not go into the LLM, tracking, tracing, or any memory DB. Define a framework with different data types and their severity levels for PII/secret detection; if severity is above 90 percent, block and redact the secrets immediately so they never reach LLM models, and blocked secrets must not go to the memory DB, tracking and eval tables, or tracing systems either. Block sensitive data at the moment of ingestion — the moment it is typed.
AI Engineer · Paving the Road to Production: Governed Enterprise AI Agents — Harish Gaggar, Credit Karma
Recommendation · MoneyHard environmental separation could have stopped the Lemkin incident: if the agent had been executing in an isolated remote workspace without a route to production, it doesn't matter what the prompt said — the agent would have had no path to do damage. Short-lived scoped credentials could also have stopped it: the agent got access to production because it found a credential in a forgotten file that had standing write access and was long-lived. A credential that can expire on its own or only carries the scope of the task itself makes it impossible to make a delete with the equivalent of an ID badge that lets you do everything. The last important piece is having audit logging that you own, not the agent: in this case the agent was the one writing its own record of its actions, which is how
AI Engineer · Your AI Isn’t the Problem. Your Infrastructure Is. — Jiachen Jiang, Coder
Recommendation · MoneyPaden Gayle recommends starting with the least authority that solves the problem, then asking what tasks can be proven with evidence to grant the system more; granting more authority than the problem requires creates uncontrolled operational risk.
AI Engineer · Deploying AI Inside a Bank Is Not an Engineering Problem — Paden Gayle, Applied AI Engineers
Recommendation · MoneyPaden Gayle recommends asking 'Who owns the source?', 'Who can grant access?', and 'Who can approve using this in the workflow?' when confirming data ownership, and warns that building AI infrastructure without understanding which datasets can be used and who can approve that use leads to deployment failure.
AI Engineer · Deploying AI Inside a Bank Is Not an Engineering Problem — Paden Gayle, Applied AI Engineers
Recommendation · MoneyPaden Gayle warns that building an AI system without a credible path of approval means it will not ship regardless of technical quality; the MVP should be built with that approval path in mind.
AI Engineer · Deploying AI Inside a Bank Is Not an Engineering Problem — Paden Gayle, Applied AI Engineers
Recommendation · MoneyUse policies and criteria to decide when it is appropriate to approve changes: auto-approve dependency bumps, documentation changes, non-breaking security patches, and lower-risk internal tooling changes; add custom approval-gating rules written in plain English; but require a human approver on top for changes touching sensitive areas like payments or auth.
AI Engineer · Rebuilding CI for AI-Scale Commit Volume — Gautam Korlam, Sonar
Recommendation · BodyDr. Andy Galpin says he prefers doing something more like three days a week of total body for most people — not because it's more effective, but because it's more resilient to life disruptions. With three days a week total body, if something happens, you've only missed that body part for 48 to 72 hours. He notes body part splits are okay for a lifting junkie who is very consistent in their schedule, but most people are not that. He also suggests that doing two days a week of whole body plus two days a week of a body part split hedges against all risks, as long as you hit the total volume number.
Andrew Huberman · Essentials: Optimal Protocols to Build Strength & Grow Muscles | Dr. Andy Galpin
Recommendation · BodyOutside of not training at all, the number one mistake people make in pursuit of fitness goals is failing to progressively overload. Progressively overload your training — increasing sets, reps, and weight over time — to continue causing adaptation and see gains in strength and hypertrophy.
Andrew Huberman · Essentials: Optimal Protocols to Build Strength & Grow Muscles | Dr. Andy Galpin
Recommendation · BodyDr. Andy Galpin advises that if you're trying to grow a specific body part or muscle, it's very important that you're actually seeing progress in that area. Don't worry about what the textbook says an exercise is supposed to target — if you're not in the right position or at the right angle, it may do very little for that muscle. If an exercise isn't targeting the intended muscle, you may need to adjust to an incline bench, decline bench, or pec fly. He describes a coaching scenario where squats are excellent for quad development in general but may not work for a specific individual, in which case a coach may move them to a machine to isolate that muscle group and ensure development.
Andrew Huberman · Essentials: Optimal Protocols to Build Strength & Grow Muscles | Dr. Andy Galpin
Recommendation · MoneySteve Jamison recommends calculating the intelligence cost of a PR by combining what you pay your engineers plus what you pay for AI into one cost bucket, and taking the ratio of cost to throughput to get the answer on whether AI spend is worth it.
Attributed to Steve Jamison as his recommendation; not a system guarantee.
AI Engineer · Tokens Are the Cheap Part — Steve Jamison, Jellyfish
Recommendation · MoneySomya Rai's advice: use two separate models rather than one combined model, because physical and online transaction portfolios have different base rates and different signal geometry.
AI Engineer · Judge a Transaction by the Company It Keeps — Somya Rai & MD Sharique, EXL
Recommendation · MoneyMD Sharique's advice: concatenate three order-invariant statistics (mean, max, standard deviation) rather than just averaging. A single dangerous neighbor among many genuine ones is exactly the pattern fraud rings create, and mean pooling is precisely the operation that would hide it; standard deviation distinguishes a uniform, boring neighborhood from one with a live outlier.
AI Engineer · Judge a Transaction by the Company It Keeps — Somya Rai & MD Sharique, EXL
Recommendation · MoneySomya Rai's advice: optimize for precision at a fixed K (the alert slice) rather than area under the curve — a fraud model must move real fraud into the top alert slice at the same latency to be worth anything.
AI Engineer · Judge a Transaction by the Company It Keeps — Somya Rai & MD Sharique, EXL
Recommendation · MoneyMD Sharique's advice: hold a soft target fraud rate of 0.5% for card-present merchant transactions (about 10× the 0.05% natural rate) and 6% for online transactions (about 10× the 0.6% natural rate). Holding a soft target fraud rate without breaking connectivity was, per Somya Rai, the key to graph sampling.
AI Engineer · Judge a Transaction by the Company It Keeps — Somya Rai & MD Sharique, EXL
Recommendation · MoneyMD Sharique's advice: make each calendar month its own disjoint graph with no edges crossing the boundary, and fit every scalar and vocabulary on training months only, then freeze them, to prevent distribution leakage. Building graph features leakage-safe is harder than it sounds when edges carry timestamps plus 32 derived fraud-detection features.
AI Engineer · Judge a Transaction by the Company It Keeps — Somya Rai & MD Sharique, EXL
Recommendation · MoneyFor aggregation, don't just average — concatenate three order-invariant statistics: mean, max, and standard deviation, then combine with the node's own vector through learned weights. With four neighbors where three look normal (scores around 0.1) and one is 0.92, averaging gives 0.32 and dilutes the ring-linked neighbor into the crowd, but taking the max lets it come through untouched at 0.92. A single dangerous neighbor among many genuine ones is exactly the pattern fraud rings create, and mean pooling is precisely the operation that would hide it. Standard deviation adds a third view: how much the neighborhood disagrees with itself. Concatenating three statistics costs nothing extra at serving time — no additional learned parameters, and still a fixed-size output no matter how many
The four-neighbor example is hypothetical, as stated by the speaker; some figures carry a possible numeric error flag.
AI Engineer · Judge a Transaction by the Company It Keeps — Somya Rai & MD Sharique, EXL
Recommendation · MoneyMichele Grant's third artifact is the autonomy scorecard, which the steering committee uses to grant autonomy in four stages. Stage one is shadow mode: analysts build the report as they do today, the agent runs alongside, and the team compares the two. Stage two is the supervised draft: the agent compiles the first draft and the senior analyst reviews every line item. Stage three is exception and sample: the analyst reviews lagged initiatives plus a random sample of claims. Stage four is exceptions only: the analyst reviews only high-risk anomalies, supported by periodic independent sampling.
AI Engineer · The Agent Works. Does Anyone Trust It? — Michele Grant, Leading at Scale
Recommendation · MoneyMichele Grant says that for every agent, leadership has to decide four things: what it is permitted to do, who owns it, what evidence supports it, and where its authority stops. She stresses this is an executive operating mandate, not an engineering task.
AI Engineer · The Agent Works. Does Anyone Trust It? — Michele Grant, Leading at Scale
Recommendation · MoneyMichele Grant says evidence for greater agent autonomy comes from how the agent is built, and she looks for three artifacts: an auditable decision log, a validation document, and an autonomy scorecard.
AI Engineer · The Agent Works. Does Anyone Trust It? — Michele Grant, Leading at Scale
Recommendation · MoneyMichele Grant insists that every single metric requires a pre-deployment baseline. If the IMO cannot state the baseline today, do not deploy — if you can't measure where you started, you can never prove the agent moved the needle.
AI Engineer · The Agent Works. Does Anyone Trust It? — Michele Grant, Leading at Scale