Sahil Sohani
Backend Engineer
I build backend systems, AI pipelines, and distributed infrastructure. Published IEEE researcher. AdvaRisk · DRDO.
4
Projects Built
2
IEEE Publications
2
Industry Internships
Education
B.E. Computer Engineering
Dr. D.Y. Patil Institute of Technology, Pune
/projects
Featured Systems
Production software built to handle real workloads. Each project reflects deliberate engineering decisions.
InvoSure
AI-powered invoice verification system
Built an AI-powered invoice verification system using Groq LLM for vendor and GST entity extraction, automated GST verification via Playwright and 2Captcha, with a Dockerized FastAPI backend and React frontend.
PI-EVA — MA-RAG
Uncertainty-aware multi-agent RAG pipeline
Production-style Multi-Agent RAG pipeline on HotpotQA using FAISS vector retrieval, Groq LLM orchestration, and Paraphrase-Induced Epistemic Variance (PI-EVA) for uncertainty quantification. EM: 44.0 | F1: 54.29.
ResQ Vision
Real-time CCTV accident detection system
Real-time CCTV accident detection using YOLOv8. FastAPI backend, Streamlit dashboard, accident evidence clip generation, SOS alerts via Telegram, and GPT-4o Mini for contextual descriptions. Deployed on LeapSwitch via Cloudflare Tunnel.
Feedback Sentiment Analyzer
Offline NLP feedback classification for DRDO ERP
Fully offline real-time sentiment analysis system using DistilBERT. Classifies feedback into complaints, queries, suggestions, and appreciation. Integrated into DRDO internal ERP portal, reducing manual processing by over 70%.
/experience
Engineering Experience
Production internships building systems that handle real data, real scale, real pressure.
March 2025 – June 2025
Backend Development Intern
AdvaRisk, Baner, Pune
Mission
Built and maintained RESTful APIs, async task pipelines, and scraper orchestration systems for a financial risk intelligence platform.
Engineering Impact
- Developed and maintained RESTful APIs using FastAPI and SQLAlchemy for scalable web scraping and data ingestion pipelines.
- Automated VMN services and scraper orchestration, improving data reliability while reducing manual intervention.
- Implemented asynchronous task queues using Celery, RabbitMQ, and Redis to improve scraping throughput and backend performance.
- Worked on cloud deployment and virtual machines using LeapSwitch.
- Provided production support in an Agile environment using Jira.
September 2024 – December 2024
Research & Development Intern
DRDO – Defence Research & Development Organisation, Dighi, Pune
Mission
Developed a fully offline real-time feedback sentiment analysis system integrated into DRDO's internal ERP portal.
Engineering Impact
- Built a fully offline real-time feedback sentiment analysis system using DistilBERT.
- Classified textual feedback into complaints, queries, suggestions, and appreciation.
- Integrated the solution into DRDO's internal ERP portal.
- Eliminated manual category selection and star ratings, reducing processing time by over 70%.
- Developed a scalable FastAPI backend with OracleDB and a Streamlit frontend.
/research
Publications
Peer-reviewed research at the intersection of AI systems and real-world deployment. Two IEEE publications in 2026.
Uncertainty Aware Multi-Agent RAG Using Paraphrase-Induced Epistemic Variance Analysis
Designed an uncertainty-aware Multi-Agent RAG framework that quantifies epistemic uncertainty using paraphrase variance. Improves response reliability through confidence-aware reasoning on multi-hop QA benchmarks.
Key Contribution
Introduced Paraphrase-Induced Epistemic Variance (PI-EVA) as a method to detect hallucination risk in RAG outputs. Achieved Exact Match of 44.0 and F1 of 54.29 on HotpotQA dev-distractor.
InvoSure: Smart GST Invoice Verification System
Proposed an AI-driven invoice validation framework combining OCR, LLM-based entity extraction, and automated GST verification. Designed a scalable backend architecture for end-to-end invoice processing.
Key Contribution
Combined OCR text extraction, Groq LLM entity parsing, and Playwright-based automated GST portal verification into a single pipeline — eliminating manual invoice validation entirely.
/philosophy
How I Build Software
Principles that guide every architectural decision, not aspirational values posted on a wall.
Build for observability.
Logs, metrics, and traces from day one. If you cannot measure it, you cannot debug it in production.
Automate the repetitive.
Every manual step is a future outage. Infrastructure-as-code, CI/CD, and self-healing systems reduce human error.
Measure before optimizing.
Profile first. Premature optimization creates complexity without evidence of bottlenecks.
Prefer simple architectures.
A well-designed monolith often outperforms a poorly-designed microservice mesh. Complexity must be justified.
Fail gracefully.
Dead-letter queues, retries with backoff, circuit breakers. Systems fail — design so they degrade, not collapse.
Design systems that scale.
Stateless services, horizontal partitioning, async task distribution. Build for 10x before you need it.
/contact
Get in Touch
{
"email": "sahilsohani2704@gmail.com",
"github": "github.com/SahilSohani27",
"linkedin": "linkedin.com/in/sahilsohani"
}