I don't build demos — I build things that survive contact with real users. Senior Software Engineer at Innoviti Technologies, building AI systems that run under real production load. My systems process 100,000+ transactions a day, serve 50+ concurrent LLM requests, and have delivered $104K+ in quantified cost reductions.
|
Hermione — NLQ-to-SQL Analytics Engine
Schema-aware natural language to SQL on Django + Qwen3-VL 27B (8-bit) on AWS EC2. Auto-visualization layer
generates charts, graphs, and maps from PostgreSQL results.
|
Holmes AI — Image QC Pipeline
Async pipeline on FastAPI + RabbitMQ processing 5,000+ support ticket images daily. S3 pull, YOLOv9
checkbox detection, Qwen3-VL 32B semantic validation.
|
|
Link — Fraud Detection
Real-time fraud detection across a 21,000+ merchant network. RabbitMQ event ingestion, feature engineering
on transaction velocity and history, LightGBM classifier.
|
vLLM Inference Stack
Built from scratch — 96GB VRAM, tensor parallelism, 8-bit AWQ quantization, continuous batching behind an
async FastAPI gateway. 128K context window, hosted locally.
|
|
Sheldon — Text-to-Config Automation
LLM-powered engine that parses unstructured brand emails and Excel dumps into PostgreSQL-compatible
records.
|
Disha-AI — Address Normalization
Two-stage pipeline: Google Maps API geocoding + Gemini 1.5 Flash semantic correction, Redis-backed
deduplication and rate limiting.
|
|
ASO Dispatch Algorithm
Geo-optimized field engineer routing across 20,000+ stores and 120 engineers. Weighted bipartite matching,
Haversine proximity via OpenStreetMap, KPI-based priority scoring, spatial indexing on PostgreSQL.
|
LLM Inference
Backend
ML / AI
Cloud / DevOps
| Metric | Value |
|---|---|
| Transactions/day (fraud detection) | 100,000+ |
| Images/day (QC pipeline) | 5,000+ |
| NLQ-to-SQL queries/day | 1,600+ at 92% accuracy |
| Local inference throughput | 3,000+ tok/s |
| QC pipeline speedup | 180× (1,260s → 7s) |
| LEA fraud-complaint reduction | 78% |
| Field dispatch backlog reduction | 37% |
| Total quantified cost savings | $104K+ |
| Brand automation accuracy | 99.42% |
| Project | What it does |
|---|---|
| DataStory | 5-node LangGraph BI agent — NLQ-to-SQL, auto-charting, anomaly detection, plain-English narrative |
| Clause & Flaws | Contract intelligence — RAG over a legal knowledge base, risk scoring, inline PDF annotation |
| IdeaForge | Multi-agent product validation — synthetic personas, live focus-group simulation, structured reports via WebSockets |
| DebugDuck | AI debugger — LangGraph chain streams root cause → fix → regression test via SSE |
| Cortex Dashboard | AI infra monitoring — p50/p95/p99 latency, GPU memory, token cost tracking, auto post-mortems |
| InventoryWhisperer | Demand forecasting with Prophet + XGBoost, automated reorder workflow via LangGraph |
| MarketAnalytics | Autonomous market research agent — real-time web search, event timeline, competitor signals |
Not research that stays in papers. Not demos that never see production. If you're solving hard problems at scale, let's talk.
