Mashreq, Atrax, NotionBench, research, products, open source
Everything on the resume that is not Cisco. Mashreq and Atrax carry the identity questions, so they are deepest. Research and products get one line each unless asked. check marks a guess.
Mashreq Bank: AI/LLM intern, Jun to Aug 2025
Explain this in 45 seconds.
Answer
Three internal tools grounded in the bank's policy documents with retrieval: code optimization suggestions, policy checks (does this document or change comply), and document quality review. Plus an MCP server that gave the bank's AI tools one shared API, and an intern operations portal in React, FastAPI and PostgreSQL so intern tasks lived in one app instead of spreadsheets and email.How did you stop the model from inventing a policy?
Answer
check Retrieval first: find the relevant policy passages, put them in the prompt, and require the answer to cite the passage. No source, no answer; the tool says it could not find a policy instead of guessing. Confirm whether you measured this (reviewers grading outputs, a set of known questions) or whether the gate was reviewer iteration.How did you handle sensitive data at a bank?
Answer
check Say exactly what is true. Where did the models run (bank-hosted, a cloud tenant, a vendor API)? What left the bank's boundary and what was redacted? Who approved the data flow? If the honest answer is "internal deployment, nothing left the tenant, compliance approved the document set," say that. If you do not remember a detail, say what you would check. Do not improvise a control that did not exist.The MCP server: what did it expose, who could call it, and how did they prove who they were?
Answer
check One shared API over the bank's AI tools so each new client did not re-implement them. The identity question: how did a caller authenticate (per-client credentials, a shared key, network trust), what could each caller reach, and were calls logged? If the real answer is "internal network plus a shared key," own it, then say what you would do now: per-client credentials, scoped tokens per tool, an audit log of every call with the caller's identity, and rate limits. This is the most natural place in your resume to show you think about identity.The intern portal: how did login and roles work?
Answer
check Who logged in (interns, managers), how (bank SSO or local accounts), and how roles were enforced on the API, not just hidden in the UI. The interviewer is listening for "authorization checked on the server for every request."What did a code-optimization tool have to do with compliance?
Answer
The bank's standards for code were policy too: logging rules, secret handling, approved libraries. The tool suggested changes against those documents, so review time went down and reviewers spent it on judgment calls instead of checklist items.Atrax: agent-native cloud platform (atrax.run)
Explain Atrax in 30 seconds.
Answer
A coding agent declares an app in one config file; the platform puts it online with a public URL, its own SQL database, and ordered migrations, on any cloud provider. Built in TypeScript. Backed by $20K in AWS credits. The point: agents can ship full apps without a human clicking through a cloud console.How are apps isolated from each other?
Answer
check Say what exists today: separate database per app, separate credentials, separate process or container, network rules between apps. Then say what does not exist yet. An honest "each app has its own database and credentials but shares a host" beats a vague "fully isolated."Where do database credentials live, and who can read them?
Answer
check Good answer: generated per app at provision time, injected into the app's environment at runtime, never written into the config file or logs, rotatable. If the real answer today is one shared admin user or credentials in the config, own it and say the fix in one sentence. The interviewer cares more that you know the right shape than that you already built it.What happens when a migration fails halfway?
Answer
check Migrations are ordered and recorded in a migrations table; a failed step stops the sequence; every step is safe to re-run; rollback where the SQL allows it (transactional DDL on Postgres, not everywhere). Confirm what Atrax actually records and whether re-running is safe today.Who is the identity here, the agent or the human?
Answer
check Good answer: the human owns the app; the agent acts under a token scoped to that app, so it can deploy and migrate that app and nothing else, and every action is attributable to agent plus owner. If today it is one platform key for everything, say so and say that this gap is exactly why you want to work on identity. This is your strongest unprompted bridge to Agent ID.How do you stop an agent from deploying something malicious, or from reaching another tenant's database?
Answer
Least privilege per app: the app's runtime gets only its own database credentials. Allow-list what the config can declare, so an agent cannot ask for a host mount or another app's network. Audit what was deployed, by whom, from what config. Treat agent output as untrusted input, the same as user input.Why "any cloud provider"? What was hard about that?
Answer
check One config, provider adapters underneath. The hard part is the lowest common denominator: databases, DNS and TLS differ per provider, so the config describes intent (an app, a database, a URL) and each adapter maps it. Say one concrete difference you hit.What did $20K of AWS credits teach you about cost?
Answer
check One sentence: what the money went to and the one cost surprise (idle databases, egress, logs). Then stop.NotionBench: coding-agent benchmark
Explain it in 30 seconds.
Answer
Open-source (MIT) benchmark in TypeScript and Docker that compared the quality and cost of eight AI coding agent setups on 38 developer tasks, 890 runs total, graded automatically with execution tests and API checks.How did you grade automatically and keep runs comparable?
Answer
Same Docker image, same task set, same grader; the only variable was the agent setup. Each run produced pass or fail from execution tests plus API contract checks, with cost and latency recorded. Multiple runs per task because agents are nondeterministic.What surprised you?
Answer
check One real finding: a cheaper setup matching an expensive one, a setup that passed tests but broke the API contract, or a task category where every agent failed. Have the number.What would you add?
Answer
Flaky-test detection, more runs per task, and a security grader: did the agent leak a secret, make an unsafe network call, or modify files outside the task. That last one is the benchmark this team would want.Research: KV-cache compression, OCT imaging
The MiniCache work, in one breath.
Answer
Clean-room PyTorch reimplementation of a KV-cache compression paper, with tests, benchmarks, and interactive notes. One sentence of why: at long context the key-value cache, not the weights, dominates GPU memory during generation, so compressing it lets you serve longer prompts or more users per GPU.The OCT demo, in one breath.
Answer
A pretrained classifier on retinal OCT scans, Grad-CAM++ to highlight the region the model used, and a drafted report. If asked about trust: a heatmap is an explanation aid, not a diagnosis; the clinician decides, and the tool is built to show its reasoning so a human can reject it.Shipped products
PDFSpacer: what broke in production and how did you find out?
Answer
check One real incident: what users saw, how you learned about it (support mail, an error log, a drop in a metric), the fix, and what you added so you would know sooner next time. Forty seconds.Share-Pen: why did you sell?
Answer
check One sentence on the real reason (time, a better owner, a good offer). Then stop. Do not narrate the deal.TrendFlip: $30K a year reselling collectibles with AI. Tell me more.
Answer
Only if asked. Used AI tools to find underpriced items and set prices; the lesson is about data and margins, not about collectibles. Keep it to two sentences; this is not the job.Published developer tools with 4,700+ downloads: what and why?
Answer
check Name the one or two that matter, what problem each solved, and what you learned from maintaining something strangers depend on (versioning, breaking changes, issues).Open source
Name projects precisely: "Kueue, a Kubernetes SIG project," "AIBrix, under the vLLM project." Not "Kubernetes core" or "vLLM itself." If you cannot explain a PR line by line, do not bring it up. If asked whether you used AI: yes, and you wrote the failing tests and can explain every line.
Tell me about an open-source bug you fixed.
Answer
Lead with Kueue. Kueue is Kubernetes job queueing: it decides when batch and AI jobs are admitted and enforces quotas. Its topology-aware scheduling records nodes by the hostname label, but the node-failure controller looked nodes up by the Node object's name. Where those differ, which is common on clouds, the lookup misses: a healthy node gets marked unhealthy, and a really failed node never matches, so its workloads are never evicted. Fix: key everything by hostname, add a field index from hostname to Node, fall back to the name when the label is missing, and if two nodes share a hostname, log an error and skip, because retrying cannot fix it. Wrote the failing tests first, then the fix, then ran the full suite. The lesson, two identifiers for one thing is a production bug waiting to happen, is an identity bug in miniature. Say that.Another one?
Answer
The llm-d router kept a cache of which serving pod held which prompt prefix. A restarted pod came back at the same address with an empty cache and its sequence counter at zero, but the code never checked the sequence, so the router sent requests to a pod whose cache was gone. Fix: track the sequence per publisher on the live path; going backwards means a restart, so clear that pod's state. Lesson: use sequence numbers or epochs to detect restarts and drop stale state. Also an identity-shaped bug: same address, different instance.What is a reconcile loop?
Answer
A controller compares desired state with actual state and acts to converge them. It is level-triggered, not edge-triggered, and must be idempotent, so running it again is always safe. Most of the bugs you fixed were places where that idempotency or identity assumption broke.How do you approach a bug in a codebase you have never seen?
Answer
Read the issue, reproduce it with a failing test, read the code path the test exercises, fix the root cause not the symptom, show red to green, run the full suite and lint, and write the PR description so a maintainer can verify it in two minutes.Leadership and programs
OpenAI Campus Lead: what did you actually do?
Answer
One of 140 students in the first cohort. As president of Penn State's OpenAI Student Collective you ran workshops, studio hours and showcases teaching ChatGPT and AI coding tools. If asked for a lesson: teaching people to use agents well taught you where they go wrong, which is where you started caring about what an agent should and should not be allowed to do.The Nas Company residency?
Answer
One of 11 from 18,000 applicants. Built Briefly, a newsletter product for creators, from user research to launch in a few weeks. Lesson: talking to users before building. One sentence, then stop.Hackathons and Jane Street puzzles?
Answer
Five hackathon wins, including a $10K individual prize for a housing project. Five Jane Street puzzles solved. Only if asked; one sentence about what the housing project did.Homelab and self-hosting: what do you run?