Critique-driven questions: imprecise percentages, skills listed versus evidenced, GitHub mismatches, every bullet in twenty seconds. Job 10454435 still none. Unnamed stays unnamed. R1/R2 done. This book is for R3 and R4. Source: Resume_Adarsh_Vishwkarma_Aug26.pdf. Labels: Resume-derived IE-asked.
A Bar Raiser who walks the resume is testing Earn Trust, not vocabulary. Agree with the imprecision, then the mechanism. Do not invent Qdrant, Haar, missed ERC, missed 99.9 percent, mentee count, p95, TTL, AWS bill, or student headcount. If GitHub contradicts the PDF, the resume wins in one sentence. Full mock Qs: _answers/08-resume-mock.md.
The PDF uses 40%, 99.9%, 65%, 35%, 60%, 210 ms, 200+, 17th/80+, 40%, 55%, 30%, 50+, 35%, 50%, 2.5×, 3–10 min, 73→89, 24 FPS, 50%, 2×. None of those come with a methods appendix. Honest answer: resume number plus the mechanism, plus what you would add. Do not invent a new decimal.
Faster than the path before query and serving work on the same BI app: report generation on Postgres plus the deployed path, not a second product. I will not invent a before-and-after wall-clock or a p95. I will defend 40 percent as the resume claim next to 99.9 percent uptime. Mechanism: SQL and deploy work, Redis on the bot path as a sibling metric (minus 35 percent database latency), dashboards off the LLM so tiles stay sub-210 milliseconds. If I measured again I would publish a before/after on the same report set. I did not write that protocol on the PDF.
Do not steal IQVIA “200+” into a percent. Do not say we missed 99.9 percent in the before period.
The resume states 99.9 percent uptime on the BI app. I will not invent a month, an SLA contract, or a missed month. I will not invent an AWS bill for the redundancy. Serving path: CloudFront, ECS, Docker, CI/CD, ALB. Highest Standards backup: BI that is down is not AI. If they want how I would monitor: health checks on tasks, ALB stopping traffic to a bad task. I will not fake a PagerDuty rotation I did not have as intern.
For people who were not writing SQL. The chatbot is LangChain SQL RAG behind three-tier RLS. Productivity is they could ask in English instead of waiting on an analyst. I will not invent a headcount of analysts or a time-and-motion study. Mechanism: NL to SQL as the user role, Redis minus 35 percent so it was usable. Isolation is part of the number: a leaky bot is not productivity.
The resume number is minus 35 percent bot database latency via caching. I will not invent p95 or TTL. Uncached every NL turn hit Postgres. Cache schema and repeats, still tier-scoped. That is Dive Deep. Differently: denied-tier tests from day one; write invalidation rules down with RLS. The number I defend stays minus 35 percent.
Thirty real-time KPI dashboards with React and Recharts, plus 60 percent operational efficiency as on the resume. Operators got tiles instead of ad-hoc asks for the same thirty. Sub-210 milliseconds, no LLM on that path. I will not invent which thirty metrics by name if the PDF does not list them. I will not invent a socket bus.
Dashboard serving path: CloudFront, ECS, Docker, the KPI APIs. Not the chatbot’s generation time. Not an LLM round trip. If they mix 210 ms with GiftedBooks sub-300 ms: different products, different numbers. I keep them in their lanes.
Because the resume has no extra IQVIA percentage. I will not add one. Ranked more than two hundred sites. Processed more than two-hundred-page BRDs with traces and test-case generation. Quality is LangSmith fail-if-no-span, not a made-up accuracy. If they want 10×: more ranker, not more Playwright on every URL.
Team score. Core software team member. I use I for GStreamer, ZED path, costmap work I owned. I use we for 17th. Obstacle plus 40 percent and collision minus 55 percent sit on the pipeline, not on a solo medal. I will not invent a missed ERC to make 17th sound like recovery.
Resume: YOLOv9 89 percent mAP improved from 73 percent, 15,000+ images, 24 FPS. I will not invent a split, precision, or recall beyond mAP. I would lock a held-out camera earlier so 89 percent is not leak — would, not published. 73 percent was a fail for PPE. Then twenty-plus cameras, violations minus 50 percent, compliance 2×, after the gate.
Resume: 2.5× faster content comprehension; doubts from hours to 3–10 minutes; plus 35 percent reading; plus 50 percent engagement; sub-300 ms; 99.5 percent uptime. I will not invent student headcount. Mechanism: RAG on their PDF, PYQ as analytics not an LLM syllabus. GitHub currently mismatches AegisAI. Resume wins.
Browser-based inference dropped iteration 30 percent versus a local loop. Marketplace more than fifty pretrained models and datasets. I will not add a user count or TAM. GitHub README is default Next.js. Resume features win. Do not merge Horizon’s field rover.
The skills block is wide. Intern bullets are narrow. A Bar Raiser will pick a skill that is not in a bullet. Say: skill versus shipped. Do not fake ops.
| On skills list | Evidenced on a bullet | Speak |
|---|---|---|
| Java | Live Code language; not intern runtime | Loop is Java; Ylogx/IQVIA Python+TS; Horizon ROS2 |
| Kubernetes | Ylogx: Docker, ECS, CloudFront | Skill; I was not cluster owner; no fake k8s ops |
| GraphQL, ProtoBuf, WebSockets | Ylogx REST + Recharts KPIs | This BI was REST/JSON; describe polling vs WS; do not invent a bus |
| MongoDB, Firebase, Supabase, MySQL | Ylogx PostgreSQL; Argus PostgreSQL | Postgres for RLS and events; others are skills not those warehouses |
| Vector DB (FAISS) | IQVIA: Azure AI Search; GiftedBooks: RAG unnamed store | Do not rename Azure to FAISS. Do not say Qdrant |
| Graph DB (Neo4j) | IQVIA: GraphDB hops | Graph hops on BRDs; do not invent Neo4j cluster ops |
| TensorFlow | Argus: YOLOv9, OpenCV | YOLO path is the evidenced CV; do not invent TF training logs |
| Local LLMs, OpenAI API | IQVIA LangGraph; Ylogx LangChain SQL RAG; GiftedBooks RAG; Stratify Gemini | Name the product; do not dump every brand |
| OAuth 2.0, JWT, Auth0, RBAC, RLS | Ylogx RLS+RBAC 3 tiers | Tokens at API; rows in Postgres. Not IQVIA multi-tenant fiction |
| Nginx, Prisma, Redux, Three.js, ShadCN | Not named in intern bullets | Skill; do not invent Ylogx as Prisma/Nginx story |
| GitHub Actions, CI/CD | Ylogx automated CI/CD, Docker, ECS | Image build, push, task roll — not a twenty-stage lecture |
If they point at Kubernetes: the defensible Ylogx path is Docker plus ECS plus CloudFront plus CI/CD. I will not fake kube-proxy. If they point at GraphQL: thirty KPIs did not need it first. If they point at Kafka: I did not operate Kafka. Adjacent is Redis plus async jobs as CS, not a resume claim. If they point at Java on IQVIA: I did not ship that intern as a JVM service.
igreaper-style Kafka / B versus B+ is Standard CS. Fat nodes for disk; B+ values in leaves for range scans. Say as CS, not “I tuned RDS.”
GiftedBooks. The giftedbooks README currently describes AegisAI. That is a mismatch. GiftedBooks answers come from the resume only: VR labs, avatars, PDF RAG, PYQ, sub-300 ms, 99.5 percent, plus 35 / plus 50 / 2.5× / 3–10 minutes. I will not mix AegisAI features or metrics. If they have the tab open: one sentence, then the student story.
StratifyLabs. GitHub is a default Next.js README. Resume wins: 3D lab, browser inference minus 30 percent, marketplace 50+, URDF/WebGL, Gemini RAG bots in the scene. I will not invent README features. I will not merge Horizon Gazebo into this browser lab. karyanode as a GitHub module is not a Stratify resume metric unless they found it — then label GitHub in the first sentence.
argus-stream-api-server. README 404. I will not invent routes. Resume: YOLOv9, 73 to 89 mAP, 15k images, 24 FPS, 20+ cameras, Postgres logs, minus 50 percent violations, 2× compliance. Video off the database. Events in.
Gstreamer-UDP. Supporting webcam-stream artifact for the Horizon 60 FPS camera story. Not a second resume project. Not extra metrics. Not the internship title.
InstaRecon / PhiSiFi. Not a project pitch. Security-awareness demo, consent, no production attacks. Then StratifyLabs, Argus, Ylogx, or IQVIA. No phishing, credential, or exploit steps. Not an LP story.
ConvBI / Warpflow / valAgent. Not Aug 2026 resume bullets. If they found the repo, label GitHub in the first sentence. Do not attach Ylogx 99.9 percent or IQVIA 200+ to those names. valAgent SecretStr is not the IQVIA security model. Ylogx RLS is.
Earn Trust is saying the mismatch out loud before they do. Pretending the README is the intern is how you fail the artifact test.
Rehearse these. Then stop. Let them dive. “Walk me through your resume” is still IQVIA plus one number from Ylogx or Horizon — not this entire list in one breath. Full Q → spoken → trap: _answers/08-resume-mock.md.
LangGraph Deep Research: planner, parallel Firecrawl, Bing, DuckDuckGo, Playwright, rank more than two hundred websites on quality, recency, agreement. Failed scrape does not restart the run. No extra percent.
Azure AI Search hybrid plus semantic, GraphDB hops, LangGraph adaptive retrieval on more than two-hundred-page BRDs. LangSmith evals and test-case generation. Cite or abstain. Not Qdrant.
Full-stack AI BI: FastAPI, NestJS, Postgres. Forty percent faster report generation, 99.9 percent uptime. I interned November 2024 to October 2025.
LangChain SQL RAG, plus 65 percent analysis productivity. RLS and RBAC for three organizational tiers. Redis minus 35 percent bot database latency. SQL as the user role.
Thirty KPI dashboards, plus 60 percent ops, CloudFront, ECS, Docker, CI/CD, sub-210 milliseconds, GoDaddy with Route 53 through an ALB. No LLM on that path.
Semi-autonomous Mars rover, ROS2, ERC 2024, 17th of more than eighty. GStreamer 60 FPS. Obstacle detection plus 40 percent. Core software, February to June 2024.
ZED 2 at more than two million points per second into RViz and Gazebo. Costmap fusion, collision risk minus 55 percent. Stale occupancy was the ship-blocker. No missed ERC.
CUSAT CSE, October 2022 to May 2026, CGPA 8.42. Not an LP metric. Java is in the skills list; the degree is Computer Science.
Eight national and regional. Named: CodeRecet first, MLH Best Project, Magnathon 2.0 runner-up. Rapid ship. I will not invent the other five titles.
Software team member, Team Horizon. Tech Team, IEDC CUSAT. Not a manager. No mentee count.
CV SaaS, 3D lab, browser inference minus 30 percent, marketplace 50+, URDF/WebGL, Gemini RAG bots in the scene. README is default Next.js. Resume wins.
VR labs, avatars, RAG on the student’s PDF, PYQ topics, sub-300 ms, 99.5 percent uptime, plus 35 reading, plus 50 engagement, 2.5× comprehension, doubts hours to 3–10 minutes. Not AegisAI.
YOLOv9, 73 to 89 percent mAP, 15,000+ images, 24 FPS, PPE and attendance, 20+ cameras, Postgres logs, violations minus 50 percent, compliance 2×. No Haar. README 404, no invented routes.
Current role IE-asked GFG 2025 BR. IQVIA two systems, 200+ and 200+ pages, then stop. Prince three one-liners: IQVIA 200+; Ylogx 99.9 percent; Horizon 17th — then go deep on one. I do not list six products.
Completed a project on your own IE-asked LC 7563011 R2 family. Internships were team. Pick Argus, GiftedBooks, or StratifyLabs. Problem, architecture, one hard part, one resume metric.
Imprecise “AI” everywhere. IQVIA is two graphs plus evals. Ylogx chatbot is SQL behind RLS; dashboards are not an LLM. Argus is boxes at 24 FPS. GiftedBooks is grounded PDF Q&A. Horizon is not a model. If I say AI five times I have said nothing.
Dates. IQVIA April 2026–present. Ylogx November 2024–October 2025. Horizon February–June 2024. CUSAT October 2022–May 2026. R1 was 13 Aug 2026. R2 was 18 Aug 2026. Do not use older dumps that say Round 2 = 14 Aug.
Do not clean a bullet by adding a number. Do not clean GitHub by describing AegisAI as GiftedBooks. Do not clean Kubernetes by claiming EKS. Do not clean Hire and Develop with a mentee count. Do not clean deadline with a miss. Practice-not-yours: chapter 13.