LLM serving runtimes · GenAI evaluation pipelines · cloud-backed data and backend systems
I build the parts of AI products that have to hold up after the demo: LLM serving benchmarks, guarded GenAI workflows, backend data services, deployment logs, and evaluation loops. The robotics thread in my background is now focused on perception work: turning culvert inspection footage into defect masks, polygons, and reviewable model outputs.
LLM serving runtime on GCP NVIDIA L4 with paged KV cache, continuous batching, prefix caching, CUDA Graph decode, and benchmark instrumentation.      
|
Controller-led multilingual mental-health GenAI system for English, Hindi, and Hinglish PHQ-9/GAD-7 item-level assessment with evidence extraction and safety routing.     
|
Incident response copilot using Next.js, FastAPI, vLLM, telemetry grounding, runbook retrieval, remediation gating, and analyst feedback loops.   
|
Cloud-native decision engine for supply operations using Vertex AI Search, conversational APIs, BigQuery pipelines, operational traces, and cost attribution.    
|
Distributed Robotics and Networked Embedded Sensing (DRONES) Lab — Research Aide
- Built the ML side of a culvert-inspection workflow: sampled keyframes from inspection video, used SAM-3 promptable segmentation for crack/spalling/void masks, and exported JSON masks and polygons for downstream review.
- Packaged PyTorch inference behind an endpoint-style workflow with frame inputs, model-versioned artifacts, endpoint logs, and per-run IoU/F1 reports; measured about 0.74 mask IoU and 2.3s p95 keyframe inference.
- Robotics context: the segmentation output was designed for culvert/robot inspection runs, but the core contribution was perception, model serving, and evaluation rather than navigation or controls.
Tata Elxsi — Software Engineer Intern
- Worked on the data and evaluation side of ADAS perception validation, building Python pipelines that paired scenario metadata with synchronized RGB, depth, semantic-label, pose, and calibration outputs from simulator runs.
- Converted raw run logs into replayable perception-evaluation sets across 120+ scenario tests, including lane, object, lighting, and weather variations.
- Added checks for missing frames, timestamp drift, label mismatch, and calibration gaps so bad scenario runs failed early instead of producing unreliable model-evaluation reports.
LLMate.ai — Backend Engineer Intern
- Built backend pieces for a governed natural-language analytics workflow over 50,000+ structured records: schema retrieval, parser validation, read-only query enforcement, dry-run checks, result caching, and replayable evaluation cases.
- Separated user-facing requests from slower LLM/database work with asynchronous task execution, status tracking, retry paths, and failure logs, which made long-running report generation easier to debug.
- Kept the LLM layer grounded in table metadata and query policies instead of a plain prompt wrapper, so user questions produced auditable SQL and bounded outputs.
- Languages: Python, C++, Java, SQL
- AI systems: LLM inference, model serving, vLLM, QLoRA, RLHF-style alignment, CUDA Graphs, model evaluation
- Backend and cloud: FastAPI, Spring Boot, Docker, GCP, Cloud Run, Hugging Face Endpoints, BigQuery, observability, CI/CD
- Data and perception: SQL validation, RAG, schema checks, JSON outputs, computer vision, segmentation, perception evaluation
I am looking for software engineering, ML engineering, backend/platform, and applied AI roles where the work is close to shipped systems: runtime behavior, service reliability, data contracts, evaluation quality, and production-facing demos.