7. GenAI fluency and later-slot CS

I start from the pasted prompt. If it is a tree, I write the tree. If it is “where should we not use this,” I name walls, then wait.

Brute, and why it fails. Parent map forbidden; BFS from every node is O(n²).

Optimal idea. If target is distance L down one child, this node is L+1; collect the other child at k−L−2. This is a tree walk, not RAG.

public List<Integer> distanceK(TreeNode root, TreeNode target, int k) {
  List<Integer> ans = new ArrayList<>();
  find(root, target, k, ans);
  return ans;
}
private int find(TreeNode node, TreeNode target, int k, List<Integer> ans) {
  if (node == null) return -1;
  if (node == target) { collectDown(node, k, ans); return 0; }
  int L = find(node.left, target, k, ans);
  if (L >= 0) {
    if (L + 1 == k) ans.add(node.val);
    collectDown(node.right, k - L - 2, ans);
    return L + 1;
  }
  int R = find(node.right, target, k, ans);
  if (R >= 0) {
    if (R + 1 == k) ans.add(node.val);
    collectDown(node.left, k - R - 2, ans);
    return R + 1;
  }
  return -1;
}
private void collectDown(TreeNode node, int dist, List<Integer> ans) {
  if (node == null || dist < 0) return;
  if (dist == 0) { ans.add(node.val); return; }
  collectDown(node.left, dist - 1, ans);
  collectDown(node.right, dist - 1, ans);
}

TC / SC. O(n) / O(n) worst.

I use a model where the engine of truth is already in place. At IQVIA, Deep Research plans, fans out tools, ranks, and synthesizes only cited spans or abstains — research ranked across 200+ websites. Hybrid RAG synthesizes from Azure AI Search Hybrid + Semantic plus GraphDB hops on 200+ page BRDs, with LangSmith. At Ylogx, LangChain SQL RAG may propose constrained NL → SQL only; Postgres executes as the caller role and RLS still filters. GiftedBooks answers from that student’s PDF. StratifyLabs may explain a scene; the scene graph is not the model. GraphDB stores requirement dependencies. LangGraph is IQVIA workflow state. Sharing the word “graph” is how this answer dies.

Use the model where a human still checks a citation or a SQL shape. Do not use it where a wrong output is a tenant leak, a missed helmet, a stale costmap, or a made-up total.

1. Authorization (Ylogx RLS / roles / cache identity). The model does not author row policy, assign the three organizational tiers, approve JOINs, own DB credentials, run as table owner, execute writes from NL, decide whether a Redis entry is authorized, or build the isolation key. Cached “top SKU” without tier in the key is a leak, not a −35% win. −35% is bot DB latency. sub-210 ms dashboards have no LLM.
2. DROP tables and other writes. SQL RAG returns inspectable SQL executed as the caller’s non-owner role. Catalog and writes are rejected. NL that would DROP, DELETE, or GRANT does not run. Invented unpaid totals from dumping 90 days of warehouse rows into the model are the failure mode the intern BI system exists to prevent. Postgres is the calculation engine.
3. Horizon costmap / actuation / packets. No LLM on occupancy, motor commands, fusion, time sync, or packet recovery. 60 FPS is GStreamer camera infrastructure. Stale video is not current occupancy. 17th / 80+ is a team ERC result.
4. Argus PPE alerts without a gate. The model does not generate PPE boxes, replace YOLOv9, relabel 73% mAP as acceptable because a caption looks compliant, be the official event record, or authorize alerts. 73% is a fail for alerting (misses + junk boxes). Ship bar is 89% mAP + 24 FPS, then 20+ cameras. Postgres stores events, not JPEG blobs. −50% / 2× are ops after alerts, not mAP.

At IQVIA I check expected citation IDs / overlapping span in retained chunks, and LangSmith traces node order. An empty scrape must not reach the writer as fact. At Ylogx I parse and reject catalog and writes, allow-list tables, execute as the caller role with RLS still on, and inspect SQL plus rows. Redis key has tier. Argus boxes are held-out mAP, not a caption that “looks compliant.” GiftedBooks answers stay grounded in that file.

I treat generated code as untrusted. I check compilation, a dry run on their example, complexity, and whether it smuggles a parent map or an extra array they forbade. If they pasted DSA, the verification is the Java I write this hour. A model that emits a monotonic stack still needs IO restated, O(n²) brute named, invariant typed, dry-run [2,1,2,4,3], and O(n) spoken. The model is a draft, not the interview.

I would use a model to search logs and draft a hypothesis. I would not let it restart ECS, approve a JOIN, DROP a table, or declare the costmap current. One hour means sequenced risk, not a generated runbook I have not read. Dashboard 5xx vs chatbot “thinking” vs RLS 403 vs CloudFront 403 are different classes. Resume 99.9% uptime held. Failure story on this resume is Argus 73% mAP, not a missed SLA.

I do not have the exact wording from those write-ups. Would you like architecture, evaluation, a where-not example, or the coding problem on the screen?

A process is the isolation boundary I actually shipped: ECS/Docker tasks at Ylogx, containerized Argus workers, Horizon ROS2 nodes talking over DDS. A thread is cheaper context switch and a shared heap — own stack, registers, PC. A native crash inside one shared process kills the siblings. Argus: 20+ containerized camera workers at 24 FPS after the 89% mAP gate. Horizon camera pipeline is a separate node from the costmap node. Ylogx 99.9% is process/container uptime on ECS tasks, not “I tuned JVM threads.” I do not invent a Java Thread per camera in production. Internships did not run on the JVM.

A transaction is a bundle that must look atomic. ACID: Atomic (all or none). Consistent (FK / CHECK / RLS still hold). Isolated (two builders do not each think they won dashboard.spec). Durable (WAL). A database deadlock is a cycle on row locks; Postgres detects a wait-for cycle, aborts one (40P01), the app retries with ids sorted. SQL RAG and the report builder share the warehouse; the bot does not use a table-owner DSN. I did not ship a custom detector and I did not run Bankers on the BI warehouse. 99.9% uptime is serving/deploy, not WAL fsync as a second SLA. Do not hold a row lock while the bot calls an LLM.

TCP is a reliable, ordered byte stream with a handshake, ports, ACK and retransmission. HTTP is request/response on top: methods, paths, headers, JSON bodies for Ylogx reports and dashboards. Ylogx REST is HTTP over TLS over TCP so a report row is not dropped mid-transfer. Dashboard GET has no LLM; sub-210 ms is that path. SQL RAG control plane is the same stack with a different verb. Horizon’s operator camera is 60 FPS GStreamer — if they ask that path, it is live video, not REST. Costmap occupancy is a ROS2 path, not an HTTP resource.

Live video treats a late frame as the wrong frame. I would rather drop a datagram than block the 60 FPS path on retransmission. UDP has no handshake and no delivery/order guarantee. TCP’s failure mode is head-of-line stall. Ylogx REST/TLS wants TCP so a report row is not dropped. I will not claim 120 FPS, WebRTC, or a radio BER. Costmap occupancy is a separate ROS2 path — no LLM on it, no YOLO-on-rover resume claim. Public Gstreamer-UDP is a supporting analogue, not a second project.

Bankers is deadlock avoidance: allocate only if the resulting state is safe (exists a finish sequence). Unsafe is not deadlock yet. Tables: Max, Allocation, Need = Max−Allocation, Available. Find a process with Need ≤ Available, pretend it finishes, add its Allocation back, repeat. If you get through all, the state is safe. I did not implement it at Ylogx, IQVIA, or Horizon. Bounded pools (connections, worker count) are how those systems cap resources. Ylogx isolation is Postgres RLS plus RBAC for three tiers, not a bankers matrix on SKUs.

If they still ask: order in Kafka is per partition, not global. B+ range scan is why time-ordered tenant logs want a tree index, not a hash. Then stop. I did not implement a B+. Ylogx is one Postgres. I did not shard. I did not operate Kafka. Kubernetes is skills-listed; Ylogx prod is ECS + Docker.

I can defend Docker images, ECS desired count, ALB health, and REST JSON for reports and dashboards. Docker images on ECS tasks, not Compose-as-prod, not EKS-as-shipped. REST: NestJS report/dashboard JSON; FastAPI SQL RAG. Private tenant JSON is not a CloudFront object. S3 is objects (if they ask) — not RLS rows. NoSQL on this resume is Redis cache-aside on the bot path (−35% bot DB), not a Mongo warehouse. I did not shard. I will not claim I operated the S3 copy API in production.