Meet Gajjar

• NIH U54 Lung Cancer Screening Study (SDSU JSBCAI research collaboration with Apple and Sharp HealthCare): building paired iOS + Apple Watch research app on Apple’s ResearchKit, HealthKit, SensorKit with FHIR-based EHR linkage across 6 passive sensor streams (heart rate, accelerometry, activity, sleep, surveys, e-consent) mapped to FHIR Patient, Observation, Encounter, Condition resources. Developing the ML pipeline to detect smoking behavior from wearable sensor signals feeding LDCT screening outcomes. • Built autoresearch, an internal LLMOps and AI orchestration framework: Claude Sonnet 4 (via Anthropic API + aider) iteratively edits training code under LLM-as-judge evaluation, online/offline eval pipelines, prompt management, git-based keep/revert, milestone HITL guardrails, and a FastAPI + Chart.js observability dashboard (7 REST endpoints). Ran 144 iterations / 53.5 hrs unattended on DAIC-WOZ at ∼$200 API cost; lifted dev F1 from 0.38 to 0.58 (+53% relative), beating published baselines. • Architected an agentic AI system (RAG + autonomous tool-calling) on Gemma 4 with native function calling for the ASHER humanoid robot. The LLM autonomously plans and invokes memory search, web search, and fact-persistence tools; integrated with robot hardware via MQTT (sub-50ms round-trip), running fully on-device with 1.8–2.2s round-trip response and zero cloud dependency. • Fine-tuned GPT-OSS-20B using LoRA + DeepSpeed on 2× NVIDIA A100 80GB with a novel assistant-only token training method (∼38 hrs / ∼90M tokens / 3 epochs / 25K conversation samples): 88.5% loss reduction (7.10 to 0.81) and 84% pass rate on 500-sample LLM-as-judge eval (Claude as judge) for relevance / persona-fidelity / safety vs ∼52% on base. • Developed a custom BF16 to INT4 quantization pipeline for the Perception Language Model (vision-language) + Orpheus TTS (eliminated bitsandbytes dependency, incompatible with Jetson ARM): PLM 14GB to 3.8GB, Orpheus 6GB to 1.6GB; ∼25 tokens/sec PLM inference and ∼400ms first-audio latency on NVIDIA Jetson AGX at ∼35–40W power (MAXN mode). • Built EmpathyRX, a real-time conversational AI for SDSU’s Healthcare Lab (Flask + Llama 3.1 via vLLM + Llama Guard security layer): 40–60 healthcare students across 150–200 training sessions/week, ∼1.2s p95 first-token latency with full responses streamed in 3–4s; ∼2–3% Llama Guard rejection rate (tuned for clinical-terminology false positives). • Built a multi-stream real-time computer vision pipeline for autonomous camera focus control on 2 simultaneous camera streams (wide + zoom, 1080p to 720p): per-frame YOLO object detection + 17-keypoint pose estimation + ByteTrack multi-object tracking + custom scoring engine + a PID controller (Kp=0.4, Ki=0.05, Kd=0.15) with anti-windup + hysteresis switching, running at ∼14 FPS on-device.

• Architected a multi-tenant RAG platform with citation-level attribution surfacing the exact chunk + document line, indexing ∼1.2M enterprise documents / ∼18M chunks (512-token / 50-token overlap): lifted Recall@5 by 45% (vs ada-002 baseline) via hybrid retrieval (BM25 + dense Qdrant + Cohere reranker + RRF fusion); p95 1.8s at 80K queries/day peak, serving ∼25K MAU aggregate across 7 live tenants. • Implemented advanced RAG patterns on top of base hybrid: CRAG (Corrective RAG) for retrieval-quality grading + query rewriting on low-confidence hits; Self-RAG for retrieve-on-demand + reflection-token-gated generation; hybrid parser for mixed-modality docs (text + tables + scanned images), lifting downstream accuracy ∼35% on complex multi-hop queries. Built dynamic vector DB selection per tenant (Pinecone premium / Qdrant mid / BM25 cost-sensitive). • Built a customizable multi-agent orchestration platform on LangChain, LangGraph, LlamaIndex, OpenAI Swarm, CrewAI sustaining 150–200 concurrent agent instances at peak across 10+ enterprise clients (US/UK/APAC); ∼78% task success on a 200-task eval suite (LangGraph 83% / CrewAI 70%) across Claude, GPT-4, Gemini. Median 4–6 steps per task, long tail 12–15 for multi-doc research workflows. • Delivered 5+ greenfield AI solutions in regulated industries: KYC pipeline (5-stage: image-quality → vision extraction → field validation → expiry check → Azure Document Intelligence fallback) processing 2,500 docs/day at 94% field accuracy (6 min → 45 sec/doc); banking-client sentiment-classification on 8–10K emails/day at 91% accuracy (5-class) routed to MongoDB + Power BI dashboards; AI-powered stock analytics dashboards in Next.js; banking email-response workflow on Temporal. • Productionized real-time voice AI on FastAPI WebSocket cluster + Twilio Voice + OpenAI Realtime API at 700–900ms median latency on the native speech-to-speech path with MCP-based tool execution; up to ∼50 concurrent calls in production, ∼40–50K call minutes lifetime. Also designed 20+ enterprise workflows on n8n + Zapier + Foundry across sales / marketing / support. • Engineered Temporal-based ingestion architecture (chose Temporal over FastAPI background tasks for guaranteed-once execution + retry + observability): scheduler + instant workflows + per-file isolated workers for RAG document indexing, S3 + MinIO storage, document resync feature, Google Drive / Confluence sync. Deployed containerized on-premise chatbots (Docker + nginx + Linux + code obfuscation + network isolation) for regulated clients. • Drove team to SOC 2 Type II (Security + Confidentiality) via Dependabot-driven dependency patch automation + manual alternative-library swaps + feature-flag gating + IDOR fixes + transport headers (HSTS / CSP / X-Frame-Options); integrated LLMGuard, Llama Guard, Azure Prompt Shield (input) + custom response-side scanner (output) for prompt-injection defense. Tech lead (matrix) for up to 6 engineers; supported $150–300K ARR POCs, largest contract ∼$400K ARR.

• Architected RESTful APIs on Python + FastAPI with MySQL + JWT auth for an enterprise ticketing platform serving ∼2,500 internal employees + ∼15K end users: 800–1,200 tickets/day at 80–100 RPS peak with p95 read latency under 200ms; full ticket lifecycle management + automated notifications.

AWS Academy Graduate: Cloud Architecting

AWS Academy