06 · Behavioral
Your Story Bank — verified facts, ready to speak
Every number here is from your master profile. Never inflate; the honest version IS the impressive version. Each story: STAR-shaped (Situation → Task → Action → Result), 60–90 seconds spoken.
§STAR in 90 seconds
Budget the clock before you open your mouth. A 90-second story is not four equal parts — the Action carries it.
~15s. One or two sentences of context. What was broken or at stake?
~10s. What YOU were responsible for. Say "I", not "we".
~40s. The decisions you made and why. This is where the interviewer is listening.
~20s. Exact numbers, then one sentence of what changed for the team.
1Sentinel flagship
Use for: "most exciting / most proud / tell me about a project"
Situation: Production incidents and mechanical bug-fixes ate a small team's time — report → reproduce from logs → fix → review → ship took days.
Task: I wanted the loop to close itself, safely.
Action: Built Sentinel end-to-end, solo, at Mira: GitHub webhooks detect issues the moment they're filed → the agent root-causes by querying production logs over MCP (real telemetry, not guesses) → an LLM pipeline drafts the fix and opens a PR → an adversarial AI-reviewer agent critiques it before any human sees it → engineers approve/reject via a Telegram gate. Design principle: autonomous in execution, never in authority.
Result: 100+ LLM-generated PRs merged per month (visible as sentinel/issue-* branches), incident resolution ~70% faster, and the team's review load shifted from "write every fix" to "approve or redirect."
Follow-up ammo
Hardest part? Context retrieval — right code slice + right logs in front of the model. Hence the MCP log integration.
Why the reviewer gate? 100 PRs/month is only useful if engineers trust the merge button. Trust was an architecture problem, not a prompt problem.
What would you do differently? (Have YOUR honest answer ready — e.g., earlier evals/golden set. Say something real.)
1:30
2The 2M req/month LLM platform
Use for: scale, ownership, "day-to-day work"
At Mira I run the LLM-serving API layer: 2M+ requests/month behind Kong on Kubernetes — authentication, per-tenant rate limiting, and usage-based billing for 50+ enterprise clients. It's the revenue path, so correctness and uptime aren't optional. This work contributed to 40% revenue growth. Alongside it: Kafka event pipelines with batched OpenSearch ingestion serving 5,000+ concurrent users across 20+ microservices.
1:30
371s → 2s
Use for: debugging, "hardest technical problem", methodology
A customer-facing endpoint took 71 seconds. I added per-endpoint tracing, followed the time, and took it to 2 seconds. Tell it as a method story: measure first → find where the time actually goes → fix the biggest cost → verify with the same trace.
1:30
4Call-intelligence in 2 weeks
Use for: speed, ambiguity, ownership
Vague requirement → shipped a client-facing call-intelligence feature end-to-end in 2 weeks. Use for "tell me about working with incomplete specs" and any voice/AI-product discussion.
1:30
5Petlevert internship
Use for: growth arc, early scrappiness
Backend intern on a MERN pet-services platform (1,000+ users). Query optimization: API response 800ms → 300ms, +40% query performance. Use when asked about your journey or first production experience.
1:30
§Weakness / failure questions — honest, prepared
"Your experience is light for this role."
"True — ~2 years including internships. What I'd point to is what I've operated in that time: a 2M-request/month revenue-path platform and an autonomous agent system in production. I've compressed a lot of production learning into those two years, and I'm honest about what I haven't seen yet."
"A time you failed"
Pick something REAL you remember (e.g., an early Sentinel fix that was wrong and what gate/eval you added because of it). The structure: what happened → what it cost → what you changed systemically. Never a humble-brag.
"Conflict / disagreement"
Have one real instance ready; focus on how you argued from data and what you did when overruled.
§Story Matcher — drill the mapping
§Questions YOU ask (pick 3 per round)
"What does the first 90 days look like for this role — what would I own?"
"What breaks most often today?" (signals operator mindset)
"How does the team use AI tooling day-to-day?" (your Claude-Code-native workflow is a strength — this opens the door)
"What separates the people who do well here from those who don't?"
For managers: "How is work planned — and how much of the roadmap is firefighting?"