AI Engineer

AI/ML Engineer (Contract)

Stealth AI Startup, Hybrid

• Leading development of an Automated Valuation Model (AVM) from inception to its current production-ready stage on a 4-person team — estimating a home's appraised value from its address so lenders can skip physical appraisals on standard, low-risk properties.
• Own the full ML lifecycle from data to deployment: cleaned ~38k subject–comp appraisal pairs, engineered 60+ features, built live RentCast API ingestion and a condition/quality scoring service, and shipped comparable-selection and condition-based triage logic with LightGBM training and evaluation.
• Achieved 5.1% MAPE with 68% of valuations within ±5% on a production-shaped holdout using live API data only — less than half the 11.5% MAPE of the data vendor's own AVM.

Forward Deployed AI Engineer

SBL, Remote

• Served as primary technical liaison for enterprise and growth-stage clients (Emoha, Topmate, Clinik), translating requirements into deployed RevOps and workflow-automation solutions.
• Engineered multi-source lead-enrichment pipelines that replaced manual prospecting with repeatable automated workflows wired into CRMs and multi-channel outbound.
• Built integrations across CRMs, lead sources, communication channels, and operational systems, with multi-channel outbound automation and human-in-the-loop checkpoints.
• Diagnosed and resolved production issues (integration failures, CRM sync, workflow errors) across multiple concurrent client rollouts.

AI Engineer & Fractional Developer

AnternData Solutions, Bangalore, India

• Built LLM and AI-powered applications and orchestration workflows for global tech clients including TigerData (formerly Timescale, the $1B+ database company), using modern AI tooling, APIs, and retrieval (RAG) systems.
• Delivered production-ready apps, reference architectures, and POC demos with customer engineering teams, accelerating developer adoption.
• As a fractional developer, built technical demos and reference implementations and produced developer-facing content for customer platforms.

RealDataAgentBench

• Benchmark of 1,412+ scored runs across 12 frontier LLMs (GPT-5, GPT-4.1, Claude, Gemini, Grok, Llama) on 39 real data-science tasks — scoring correctness, code quality, efficiency, and statistical validity with per-run cost tracking and 95% confidence intervals.
• Surfaced the core reliability gap — models score 0.84–0.99 on correctness but 0.52–0.90 on statistical validity — and found gpt-4.1-mini statistically tied with top-ranked gpt-4.1 at 65× lower cost per task.

CostGuard

• Production proxy that intercepts LLM calls, validates responses with RDAB-derived scoring, and retries or falls back across 12 models and 5 providers — with per-call cost tracking surfacing routing decisions that cut spend 10–20× vs. GPT-4o defaults.
• Per-provider circuit breakers, a replay endpoint returning quality deltas with 95% bootstrap CIs, Prometheus/Grafana observability, rate limiting, and a six-type alerting engine over Slack and webhooks; 90+ unit tests.

RelayOps

• Production-shaped telecom support agent with deterministic access control, scoped tools, tiered intent routing, hybrid RAG with citations, and guardrails blocking invented offers and PII — validated with adversarial evals (1.000 safe-route on a 100-case suite; zero billing escapes).

Tether & LoRA fine-tuning

• Tether: a durable execution layer (checkpoint/resume, cross-provider failover) for long-running LLM agents.
• LoRA fine-tune of DeepSeek-R1-Distill-Qwen-1.5B reaching 58.2% on GSM8K with 98.8% fewer trainable parameters.

M.S. Computer Science

Sacred Heart University, Fairfield, CT

• GPA: 3.8/4.0
• Upsilon Pi Epsilon (UPE) Honor Society

B.Tech Information Technology

GMR Institute of Technology, Vizianagaram, India

• GPA: 8.30/10