Headline

Student at SIES Graduate School Of Technology

About

I build the infrastructure that makes AI agents reliable enough to act on real systems, not just answer questions. My main project is OpsPilot, its an autonomous multi-agent platform running in production on GCP. Four agents pull live context from Linear, GitHub, Notion, and Slack through their APIs, reason over what they find, and deliver a briefing on what's stuck and what needs attention. The interesting problems weren't the model calls. They were the unglamorous ones: running rules before the model so Gemini only sees cases where context actually decides the answer, routing between Flash and Pro to cut query latency from 15s to 3s, self-healing on malformed output instead of failing the run, HMAC-SHA256 verification on every integration. 145 tests across 7 layers. Designed, built, and deployed solo. Before that, 22 months at Cognitude Consultancy building production data and AI pipelines, a Whisper and GPT-4o content engine that turned corporate training videos into navigable chapter summaries, and a PDF extraction system across 300+ institutions with a reconciliation engine that validated aggregate seat capacity against individual records and surfaced discrepancies manual review had missed. Python and FastAPI are my depth. Currently looking for backend or AI engineering.

Location

Mumbai

AI stack

antigravity

claude code

codex

gemini

kimi

langgraph

Work experience

Data Specialist at Cognitude Consultancy Private Limited, Aug 2024 - Mar 2026 · 1 yr 8 mos, Multimodal AI Content Engine (Production-Deployed) Corporate Deployment: Built and shipped an automated transcription and analysis pipeline for corporate training videos, deployed to production and used on live client content reducing the time reviewers spent locating specific material within long recordings. Multimodal Integration: Integrated OpenAI Whisper and FFmpeg for audio extraction and timestamped transcription, handling variable-length recordings and inconsistent audio quality. Structured Summarization: Developed a dual-prompt GPT-4o system that segmented raw transcripts into structured, chapter-based summaries, the first prompt establishing segment boundaries, the second generating summaries within them. Interactive Analysis: Wrote a custom Python/Regex parsing layer converting unstructured LLM output into a timestamp-synced interactive frontend, closing the gap between raw model output and a usable reviewer interface. Industrial-Scale Data Acquisition & Integrity Extraction at Scale: Engineered a Python and pdfplumber pipeline ingesting admission data from 300+ engineering colleges across inconsistent government PDF formats, extracting 10+ fields per record including college codes, course codes, sanction intake, and category-wise seat quotas. Dual-Phase Ingestion: Built a two-layer extraction system capturing institutional seat capacity at the macro level and individual candidate records at the micro level such as merit rank, percentile, category, and seat type, including TFWS and EWS classifications from separate document sets with entirely different structures. Zero-Error Validation: Designed a mathematical reconciliation engine cross-validating aggregate seat capacity against individual student records, achieving 100% data parity across thousands of entries and surfacing discrepancies that manual review had missed.

Education

Bachelor of Technology, Artificial Intelligence, SIES Graduate School Of Technology