I build software, data, and AI systems that hold up in production — reliable backends, measurable pipelines, and applications that actually ship.
I'm an engineer working across AI, software, and data science, with an MS in Data Science from Stony Brook University and 2+ years of enterprise software experience at Genpact building Salesforce CRM architectures for global clients.
My focus is building systems that are measurable and production-ready — whether that's a backend service, a data pipeline, or an LLM application. I've shipped evaluation frameworks benchmarking GPT-4, Gemini, and LLaMA across normalized metrics, edge-native apps with recursive failover, real-time serverless dashboards, and autonomous multi-agent systems. I care about the parts that decide whether software survives contact with production: clear metrics, sane architecture, and reliability under load.
ARMA AI Labs — California, USA
Genpact — India
TannMann Foundation — Remote
Backend services, data pipelines, and AI systems — each with metrics to back it up.
Novel architecture internalizing error-correction patterns via LoRA adapters and high-temperature sampling (τ > 0.7) — eliminating agent reflection collapse, cutting latency 94% (45.2s → 2.8s on Groq LPUs), and beating Reflexion baselines on HumanEval & GSM8K.
Edge-native interview platform powered by Cloudflare Python Workers (Pyodide WASM), Cloudflare D1 distributed SQL, and a Gemini → Groq → Mistral recursive failover chain. 40ms TTFB, sub-100ms cold starts, and 99.9% uptime.
Stateful multi-agent CRM orchestrator built with LangGraph, FastAPI, and React. Automates lead triage, 0-100 BANT scoring, and HITL margin guardrails with real-time Langfuse tracing and 7 DeepEval benchmark metrics.
Multi-agent cybersecurity pipeline chaining Recon (Nmap, Shodan API), Vulnerability Scanning (NVD CVE API, CVSS), Exploit Synthesis, and Strategic Reporting powered by Groq Llama-3.1-70B with Gemini fallback and strict SSRF guardrails.
Migrated from a heavyweight Kafka + Snowflake + Airflow modern data stack to global Cloudflare Workers (V8), D1, R2, React 18, and D3.js. Dynamic edge caching cuts external API load by 90% at $0 idle compute cost.
Standardized evaluation engine for professional content rewriting benchmarking GPT-4 (96% factuality), Gemini 2.5 Flash (40x cheaper), and LLaMA 3-70B across 7 normalized metrics (BLEU, ROUGE, BERTScore, DeepEval G-Eval).
Vision-based HCI system converting hand kinematics into musical MIDI in real-time at 30 FPS (<80ms latency) using MediaPipe and an adaptive LLM tuning loop that accelerates user learning curves by 40%.