Full-stack engineer focused on production ML systems, async pipelines, and LLM-powered applications. Currently building at the intersection of backend infrastructure and applied AI.
Next.js · FastAPI · PostgreSQL · Groq Llama 3.3 70B · MiniLM (local) · Celery
A multi-tenant RAG platform for querying SEC 10-K filings with cited answers, side-by-side document comparison, z-score anomaly alerts, and a live EDGAR filing watcher.
Deliberate architecture: vectors stored as JSON in PostgreSQL (zero extensions, zero second service), cosine search done in Python with NumPy — brute-force is correct at this corpus size and the escape hatch to a pgvector HNSW index is a one-column migration. Cross-encoder reranking (ms-marco-MiniLM) improves citation accuracy ~3.4 points over dense-only retrieval at ~180 ms P50 CPU cost. Local MiniLM embeddings cut $13/month off the OpenAI embedding bill. End-to-end query P95: ~1.55 s. Full stack runs at ~$386/month for 1,000 MAU vs ~$2,563 on the equivalent paid stack — a $26k/year difference driven almost entirely by Groq Llama vs GPT-4o.
Backend ships with ≥70% line+branch coverage (pytest, in-memory SQLite, CI-enforced).
React · Node/Express · FastAPI · Redis · PostgreSQL · YOLOv8 · EasyOCR · RapidFuzz
Built for HackArena 2.0. Barcode scans run through a multi-layer verification pipeline under 1 second: inventory cross-reference, scan-frequency tracking (Redis TTL counters), barcode-age analysis, fuzzy string matching via RapidFuzz, YOLO object detection, and EasyOCR label verification — producing a risk score that either approves or blocks the transaction.
Redis NX locks (txn:lock:{shopId}:{barcode}, 5s TTL) prevent duplicate processing under concurrent scans. Fraud blocks above 0.7 risk fire a SendGrid alert with an AI-generated Llama explanation. Results pushed to the checkout UI via WebSocket in real time.
Node/Express · Flask · CNN (TensorFlow/Keras) · PostgreSQL · BullMQ · Redis · Socket.io
Precision agriculture app identifying 7 potato leaf disease classes at 97% accuracy. Multi-lingual (English, Hindi, Bengali) via crowdsourced JSON i18n. Trial scan quota system (3 scans, no account required) with Google OAuth for registered users.
ML inference runs on a Python/Flask CNN backend decoupled from the Express request cycle via BullMQ + Redis — uploads return 202 immediately with a job ID, inference runs in a worker, results are pushed via Socket.io with a polling fallback. Automated diagnostic email reports via Nodemailer.
Node/Express · Flask · EasyOCR · MobileNetV2 · KNN · PostgreSQL · node-cron
Digitizes physical prescriptions via OCR (EasyOCR + MobileNetV2 region detection + KNN field classification). Extracts doctor name, medications, and next visit date. Cron-based reminder emails fire 24 hours before scheduled appointments. Personalized dashboard with prescription history and doctor stats.
Currently learning: DSA in Java · distributed systems internals
Fun fact: the first computer bug was a literal moth found in a relay at Harvard in 1947.