Open to SDE / AI Engineer roles

Pranay Lohar

I build production AI systems — LLM orchestration layers, hierarchical intent pipelines, and RAG architectures that handle real traffic. Currently shipping voice AI agents at Nurix AI.

40%
LLM reliability improvement
80%
Inference latency reduction
100+
Leads automated per day
3
Live production deployments
View ProjectsResume
Pranay Lohar
scroll
01 / experience

Software Development Engineer Intern

Bespoke Technology·Client: Nurix AI
Jan 2026 – Present
Bengaluru, India
▸

Architected end-to-end conversational AI voice agents integrating STT, LLM, and TTS pipelines across 4 production APIs (Deepgram, ElevenLabs, GPT-4.1 mini, MS Dynamics 365) for fully autonomous outbound call workflows.

▸

Designed and deployed serverless Python automation pipelines on AWS Lambda, handling pre-call lead enrichment and post-call CRM updates with sub-second trigger latency.

▸

Improved LLM agent reliability by ~40% through systematic prompt optimization and API integration testing across multi-turn voice interaction scenarios.

▸

Delivered a fully automated outreach pipeline eliminating 100% of manual sales data entry, processing 100+ outbound leads/day over 3 months.

LLM OrchestrationAWS LambdaDeepgramElevenLabsGPT-4.1CRM Automation
02 / projects
click card to expand

GeoLLM

liveAug 2025

Natural language → satellite maps + geospatial analysis

<30s
Query to map (city-scale ROI)
4
GEE analysis types
35k km²
Max supported ROI
Live DemoGitHub
Tech Stack
FastAPINext.jsSentenceTransformersFAISSRedisGoogle Earth EngineMapLibreAzure Container Apps
Problem

Satellite analysis (NDVI, LST, land-cover, water) requires GIS skills and Earth Engine code. Environmental reports live in separate PDFs. Two disjoint workflows for non-GIS analysts.

Approach

Single FastAPI monolith with 4 internal modules communicating via direct Python imports (zero inter-service HTTP). A Core LLM Agent runs hierarchical intent classification — top-level GEE vs Search, then GEE sub-classifier for NDVI/LST/LULC/Water. Results stream live via SSE with chain-of-thought steps.

Prompt2Shell

liveJul 2025

Fine-tuned Phi-3-mini that converts plain English to shell commands

30%
Shell command accuracy over baseline
80%
Inference latency reduction
3.8B
Parameter model, QLoRA fine-tuned
Live DemoGitHub

RezScan

live2025

Semantic ATS — SBERT-powered resume ranking vs job descriptions

384-dim
SBERT embedding space
2
Similarity metrics (cosine + euclidean)
1–3
Azure Container App replicas (autoscale)
Live DemoGitHub
03 / skills

LLM & Agentic Systems

core
LLM OrchestrationLangChainRAG pipelinesPrompt EngineeringIntent ClassificationSSE StreamingMulti-turn AgentsQLoRA fine-tuning

ML & Embeddings

core
SentenceTransformers (SBERT)FAISSHugging FacePyTorchScikit-learnspaCyNLTKNLP (BERT/SBERT)

Backend & APIs

core
PythonFastAPIFlaskREST APIsPostgreSQLRedisDocker

Cloud & Infra

applied
AWS LambdaAzure Container AppsModal (GPU Serverless)VercelRenderGit

Frontend

familiar
ReactNext.jsTypeScriptTailwind CSSMapLibre GL

$Depth over breadth — labeled by what I've shipped to production (core), applied in real projects (applied), or know conceptually (familiar).