I build systems that make decisions. Not just systems that process requests.

My work sits at the intersection of backend engineering, operational automation and applied AI. I have shipped LLM-powered diagnostic pipelines, computer vision inspection systems, real-time workflow engines and backend platforms that handle real operational load. Across all of it, one thing stays constant: the engineering has to be trustworthy, not just functional. Confidence thresholds, deterministic control logic, audit trails and human oversight where the stakes are high. These are not optional features. They are the baseline.

AI belongs inside a well-engineered system, not in front of it.

Every system here is built around that principle.

6 production systems  ·  Backend  ·  AI & LLMs  ·  Computer Vision  ·  Workflow Automation  ·  Decision Intelligence

Backend & APIs

Also: REST APIs · JPA/Hibernate · JavaScript · Maven

AI & Machine Learning

Also: Transformers · Pandas · NumPy · LLMs · RAG · Vector Search · LSTM/GRU

Data & Storage

Also: SQLAlchemy

Cloud & DevOps

Also: Azure · CI/CD

Dev & QA

Also: System Design · Workflow Orchestration

At 2am when production is on fire, the last thing you want is a search bar. This system already knows what broke before. It retrieves similar incidents, ranks root-cause hypotheses by confidence, decides if it is certain enough to act on its own or if it needs a human to sign off, and records every reason it made each call. Mean triage time went from 30 minutes to under 12. Unnecessary escalations dropped by 35%.

Python FastAPI LangChain Qdrant PostgreSQL Docker

The ML model gets to predict. The safety layer gets to decide. They never swap roles. That one design rule is what makes this system safe enough to control hardware while you sleep. Thermal micro-awakenings dropped by 45%. Manual temperature adjustments disappeared from over 80% of sessions.

Python PyTorch LSTM/GRU FastAPI Docker RNN

Every wrong call a computer vision model makes in a hotel is a guest complaint waiting to happen. So this system does not just classify rooms. It watches itself. Every disagreement with a human inspector gets logged, tracked and fed back into threshold calibration. 1,500 to 2,000 rooms a day. Every disagreement makes the next call sharper.

Python OpenCV PyTorch FastAPI PostgreSQL

Two incidents arrive at the same time. Two engineers grab the same one. Now one has two owners and the other has none. This engine makes that impossible. Every incident gets exactly one owner, one path through acknowledgment and escalation, with deterministic routing that holds even when multiple failures hit at once. Mean resolution time went from 22 minutes to under 8.

Java Spring Boot REST APIs PostgreSQL Docker

A 6-person hotel staff handling 52 guest requests a day sounds manageable until someone calls in sick, three rooms need cleaning at once and the front desk is running on a whiteboard. This system replaced the whiteboard. Built for ButeakSuites, it routes every request to available staff based on real-time workload, with guaranteed sub-15-second dispatch. Fulfillment time went from around 10 minutes to 4 minutes.

Java Spring Boot JPA/Hibernate Node.js PostgreSQL Docker

Bad threat data does not announce itself as bad. It just looks like data. Duplicate entries, stale records and conflicting feeds all look the same to an analyst under pressure. By the time someone realises the intelligence was wrong, a decision has already been made on it. This system sits at that point and blocks anything it cannot verify before it reaches a human.

Python FastAPI PostgreSQL Docker

Software Development Engineer Intern · BUILD91

Three companies. Six products shipped. All in one internship.

• Delivered production systems across BUILD91, EmagicorTech and Manojava Systems, spanning hotel operations, sales automation and threat intelligence.
• Took a 52-room hotel from whiteboard coordination to a deterministic workflow engine. Now the operational backbone of ButeakSuites.
• Took a sales team's daily lead capacity from under 15 to 40+ by automating every step between a lead arriving and someone owning it.
• Found 17,000 operational records silently corrupting production. Cleaned them. Reduced repeat incidents by 30%.
• Pushed task dispatch to under 15 seconds. Tripled pipeline throughput. Cut mean incident resolution from 20 to under 8 minutes.

M.Tech in Computer Science and Engineering from the International Institute of Information Technology, Naya Raipur, with a postgraduate specialisation in Data Science and AI. The statistical learning, applied ML and intelligent systems work done here is the academic foundation behind the engineering in these repos. My Undergraduate degree in Computer Science and Engineering from MIT Arts, Design and Technology University, Pune, where the software fundamentals were built.

If something in the pinned repos caught your attention, I am happy to walk through the design decisions behind it.