Hey, I'm Vanssh 👋
I'm a backend and applied-AI engineer. I care about the unglamorous half of software: the part after the demo, where a system has to survive real traffic, real users and 2am failures.
Currently an SDE intern at OpenIntervue, an AI hiring startup, working across the stack of its live AI-interview product. Outside work, I run two projects in production: CivicFix, a civic-complaint tracker on AWS, and Resolve.ai, a multi-agent system for payment collections. I'm also leading the build of Ally, an AI interviewer for Google Meet that steps aside the moment a panelist raises their hand.
Experience
- SDE InternOpenIntervueOct 2025 → Now
I work across the stack of Pihu Meet, the live AI-interview product: the AI panelist, real-time proctoring, resume parsing, and the LLM layer under all of it.
- ~790
- commits across 13 production repos
- −56%
- live LLM cost per interview
- ~70%
- lower p99 on resume parsing
- 6
- services moved to Claude
Pihu Meet: live interviews
- Built PIHU, the AI panelist that runs live interviews by voice, hands the floor to a human panelist mid-interview, and picks back up without making the candidate repeat an answer.
- Built hybrid turn-taking: a recruiter's raised hand lowers itself once the candidate starts answering, while backchannel like "mm-hmm" is ignored.
- Cut live LLM cost per notetaker interview from $0.177 to $0.078 by throttling rebuilds, merging calls and prompt caching, replayed on real interviews with quality equal or better in a blind comparison.
- Made recording survive real devices: phones record the camera (with an MP4 fallback for iPhone), laptops record the screen, a recorder that dies resumes in a new segment, and the call gets bandwidth first on weak networks.
- Made joins resilient on flaky networks: RealtimeKit retries with exponential backoff and jitter, and every client converges on the same room after a rotation.
- Shipped the lobby and meeting surface: knock-to-enter waiting room, client consent notices, background blur, mic checks, per-panelist feedback gates, and reports grounded in transcript evidence.
Proctoring
- Built the real-time proctoring service from v2 to v3: InsightFace face verification, MediaPipe head pose and gaze, YOLO object detection, anti-spoofing, mask checks and multiple-voice detection, at sub-180 ms.
- Fixed the latency problems by moving heavy inference into worker threads and loading models once per process instead of per request.
- Cut false positives with a background-object baseline at session start, minimum-confidence thresholds, lighting checks and unknown-person fixes.
- Added AI-answer detection that separates AI-generated from prepared answers, plus screen, clipboard, focus-loss and context-menu tracking rolled into a per-session compliance score.
- Rolled proctoring out to every assessment type (code, SQL, design, campus drives, technical assessments) and per interview, keyed by meeting.
Resume parsing & candidate pipeline
- Re-architected the LLM resume-parsing pipeline around streaming: ~70% lower p99 and 1,000+ concurrent parses, which unblocked enterprise batch onboarding. Prompt and model changes now pass an eval set of labelled resumes before shipping.
- Built async bulk parsing with Redis pub/sub progress tracking, and parallel scoring of every resume against the job description.
- Built parsing APIs for client integrations, with parse results and interview reports delivered by webhook and retried until delivered.
- Built OpenPool candidate matching, with candidates vectorised in Pinecone for retrieval against a JD.
LLM platform, reliability & tooling
- Moved every LLM call across 6 services from Azure OpenAI to Claude Haiku 5.5, with per-task effort, prompt caching, and cost tracked per room and per candidate.
- Found that production failure alerts had never fired: the counter was per-process across 8 workers. Moved it to a shared Redis sliding window; degradations now page the team.
- Wrote a PyTest integration suite covering 50+ API endpoints and 25+ race-condition scenarios.
- Built the recruiter Admin Dashboard end to end (API and UI): recruiter, candidate and trend views, hiring-status filters, panelist feedback states and CSV exports.
- Set up an hourly triage agent that pulls open issues from Notion, traces root cause across codebases, and writes the diagnosis back to the ticket.
Clubs & societies
- IEEE ISE · Bangalore Institute of TechnologyMember
As a member of IEEE ISE at Bangalore Institute of Technology, I conducted a hackathon for the student community.
Where I'm useful
APIs, queues and data models built for production traffic, not demo traffic.
LLM pipelines and LangGraph agents that call tools, change real state, and pass evals before they ship.
Docker, AWS, alerting. If I shipped it, I'm on call for it.
Race-condition tests, alerts that actually fire, p99s that stay flat.