Who I am
Software Engineer with hands-on experience building Gen AI systems for insurance and financial workflows. I care about turning research-grade ML ideas into reliable, auditable, high-throughput production pipelines — from document classification to retrieval-augmented generation.
2.5M+
Documents processed / year
85–90%
Straight-through processing
700+
DSA problems solved
2M+
LinkedIn views
Netaji Subhas University of Technology, Delhi
B.Tech in Computer Science & Engineering
Sant Gyaneshwar Model School, Delhi
Higher Secondary Education
Ramanjot Singh/Software Engineer · Gen AI/Delhi, India
2.5 million documents a year, classified without a human.
I build the retrieval and LLM systems behind numbers like that — and the guardrails that keep them auditable when they are wrong.
Barcode
OCR
Semantic search
LLM fallback
Each tier resolves what it can and passes the rest down. 85–90% of documents clear without a human — the remainder routes to manual review.
Where I've worked
Zinnia
Software Engineer — Gen AI (Full-Time, prev. Intern)
Jan 2025 — Jul 2026
- Architected a multi-tier document classification pipeline (barcode → OCR → semantic search → LLM fallback) processing ~8K–10K documents/day (~2.5M/year) for insurance and financial workflows.
- Achieved 85–90% straight-through processing (STP), significantly reducing manual review effort and operational costs.
- Built Power BI dashboards on pipeline logs (PostgreSQL/MongoDB) to track STP rate, throughput, fallback/error trends, and per-stage latency for real-time ops visibility.
- Built semantic classification using text embeddings + Qdrant vector search as an intelligent fallback, with LLM-based validation layers, confidence thresholds, and critic-model checks for safe, auditable outputs.
- Designed an event-driven, serverless architecture on AWS Lambda with fault-tolerant, timeout-aware, multiprocessed execution for high-throughput document workflows.
- Engineered a production-grade RAG pipeline (OpenAI embeddings + Qdrant) for Q&A over large-scale insurance document corpora, plus LLM-powered multi-document comparison for cross-policy analysis.
- Built scalable PDF ingestion pipelines with Django, FastAPI, Celery, PostgreSQL, MongoDB, Redis, and AWS S3 — chunking, embedding, and retrieval optimized for low-latency, high-precision use.
Things I've built
HireFlow AI
An AI-assisted recruiting workflow that turns a job description into structured requirements, surfaces recruiter-approved contacts, and runs consented AI screening calls.
- Built as a production-oriented monorepo with independently deployable Next.js and FastAPI containers, shipped as multi-stage images to Amazon ECR and ECS.
- Extracts structured requirements — skills, seniority, location, experience, search keywords — from raw job descriptions, every field recruiter-editable before use.
- Integrated Apollo contact search with per-job deduplication and Hunar voice agents for screening calls, persisting jobs, candidates, calls, and webhook events in Supabase PostgreSQL.
- Processes signed, idempotent webhooks for call outcomes — status, summaries, duration, recordings — surfaced live in the recruiter dashboard.
Census Insight Agent
A citation-grounded LangGraph agent over 341 pages of Census reports, built for reliable, verifiable multi-state comparisons.
- Combined Vertex AI dense embeddings, BM25 retrieval, and Qdrant reciprocal-rank fusion with per-entity evidence reservation.
- Designed an end-to-end provenance chain: PDF checksum → page-bounded chunks → verified claims → citations.
- Added one-shot citation repair, structural refusals, and an isolated, network-disabled Python executor for generating validated charts and tables.
Hybrid Deep-Learning Stock Forecasting
An advanced stock price prediction system fusing sequence models, gradient boosting, and anomaly detection over a decade of market data.
- Integrated LSTM, XGBoost, and anomaly detection (Isolation Forest, Autoencoder) over 10+ years of market, fundamental, and macroeconomic data.
- Implemented an Attention-based LSTM, cutting RMSE by up to 67.6% (Tesla), 44.9% (Amazon), and 60.6% (Nvidia).
- Achieved R² scores up to 0.99 (Apple) and 0.97 (Google, Tesla, Nvidia) — a 7% improvement in R² overall.
Tools I work with
AI / ML
Frameworks & Libraries
Databases & Search
Cloud & DevOps
Data & Visualization
Languages & Fundamentals
Highlights along the way
700+
DSA problems solved
Mastered 700+ DSA questions across LeetCode and GeeksforGeeks.
AIR 4766
JEE Mains
Achieved AIR 4766 in JEE Mains, surpassing over 1 million students.
AIR 6212
JEE Advanced
Attained AIR 6212 in JEE Advanced, among 200K qualifiers from 1 million candidates.
200+
Students placed
Served as Placement Coordinator — managed placements for 200+ students and coordinated with 10+ companies.
50+
Juniors mentored
Mentored 50+ juniors in DSA through dedicated teaching sessions.
Ask this résumé anything
Rather than describe how I build retrieval systems, here is one. Every passage of my experience is indexed and scored with BM25 and domain-aware query expansion — running entirely in your browser, no server, no API key. Ask it something.
07 — Contact
Let’s build something worth shipping.
I’m always open to conversations about Gen AI systems, interesting engineering problems, or new opportunities. Reach out — I usually reply within a day.