Skip to content

Building reliable systems for a messy, non-deterministic world

HUMAN AGENT // ONLINE

Arun Joshi

ROLE

Senior Software Engineer · Applied AI

The autonomy of an AI agent, backed by the judgment of a senior software engineer.

Give me a high-level objective, and I will navigate the ambiguity across product, engineering, and AI systems to ship it.

CONTEXT WINDOW

6+ years building production software
MS in AI
Founder experience

SKILLS.MD

Tools are easy.
Alignment is rare.

Tools and frameworks can be learned. Alignment comes from choosing and combining them to turn stakeholder vision and product requirements into reliable, trustworthy production systems.

01

Frame the System

Product & Architecture

Product StrategySystem DesignAPI DesignAuthenticationUX

02

Build the Product

Product Engineering

PythonTypeScriptNext.jsReactReact NativeC#/.NETREST APIs

03

Ground the Model

AI & Retrieval

RAGHybrid RetrievalEmbeddingsRerankingQdrantpgvector

04

Bound the Agent

Agents & Evaluation

Tool CallingContext EngineeringAI SDKEvalsLLM-as-Judge

05

Operate Reliably

Infrastructure & Delivery

AzureAWSDockerAzure DevOpsCI/CDQueuesBackground Jobs

MISSION LOGS

What I
actually built.

Receipts from the mission queue: production AI systems, full-stack products, applied ML, and the engineering behind each one.

AI SYSTEMS

Production RAG • Agents • Evaluation

Lattice

Organization-scoped AI knowledge platform with reliable ingestion, hybrid retrieval, evidence-backed answers, and a bounded agent runtime.

  • Built multi-tenant authorization, PostgreSQL/pgvector retrieval, rank fusion, reranking, context expansion, and citations.
  • Designed idempotent Celery/Redis ingestion and safe reindexing across private files, websites, and normalized content.
  • Implemented bounded tool execution with timeouts, retries, call budgets, duplicate-call handling, and SSE streaming.

AI Systems • RAG • Tool Calling

Support Agent Builder

Configurable support-agent platform adopted by five clients, with grounded answers, product tools, deployment controls, and live-agent handoff.

  • Connected knowledge retrieval, product APIs, and model-selected tools in one configurable assistant runtime.
  • Built versioned prompt and tool workflows with draft, publish, rollback, and embeddable website deployment.

PRODUCT PLATFORMS

Founder • Applied AI • Collaboration

TheDrive.ai

AI-powered file workspace for natural-language search, content generation, collaborative editing, and agentic file operations.

  • Co-founded and shaped the product as it grew to 1,000 monthly active users.
  • Built collaborative TipTap editing with socket-driven synchronization, autosave, and shared document state.
  • Designed event-triggered reindexing so edited content remained searchable by downstream AI workflows.

EdTech • Video • Distributed Work

Agile Academy

Career-learning platform supporting live instruction, self-paced courses, discussion workflows, and responsive media delivery.

  • Built live and self-paced video learning experiences with CDN-backed delivery and caching.
  • Implemented discussion workflows, queue workers, and background processing for long-running work.

Marketplace • Mobile Apps

Where Trades Go

Mobile-first workforce platform connecting contractors and skilled tradespeople through vetting, matching, and communication workflows.

  • Built mobile workflows for opportunity discovery, profiles, and employer communication.
  • Developed contractor, trade-school, and administrative portals for workforce pipelines and analytics.

APPLIED ML

DeBERTa • Multi-Task NLP • Evaluation

Emergency Dispatch Classification

Multi-task classifier trained on 6,030 real and synthetic 911 transcripts to predict incident, severity, and dispatch labels.

  • Fine-tuned DeBERTa-v3-base and achieved 0.8680 dispatch positive-label macro F1.
  • Used macro/micro F1, confusion matrices, validation loss, and error analysis to diagnose imbalance and overfitting.

Healthcare AI • Local LLMs • Privacy

Clinical De-identification

Local-first clinical workflow that removes protected health information before de-identified context reaches advanced reasoning models.

  • Fine-tuned Gemma with LoRA for local PHI removal from clinical documents.
  • Evaluated privacy and clinical-fidelity tradeoffs across rules, NER, LLMs, and human review.

WRITING & EDUCATION

Open Source • System Design • Education

AI Engineering Handbook

A sequential guide for software engineers learning reliable AI systems, from model fundamentals and hybrid retrieval to agents, evaluation, security, and production architecture.

  • Connects foundational concepts to production concerns including permissions, retries, observability, caching, and failure recovery.
  • Includes focused chapters on RAG and agent evaluation plus system-design case studies.
ARCHIVE // ADDITIONAL WORKView more projects+

Additional product and AI/ML work, kept in the page source for direct discovery and indexing.

Multi-Agent Systems • Orchestration

Consensus

Configurable AI boardroom where participants discuss, debate, vote, and produce structured recommendations from connected knowledge.

  • RAG-backed participants and knowledge sources
  • Turn-taking, voting, and meeting-control workflows

Quantization • LoRA • QLoRA

Low-Memory Model Training

Model-efficiency test bed comparing precision, 4-bit weight quantization, LoRA, and QLoRA across storage and trainable parameters.

  • Blockwise 4-bit weight packing
  • Low-rank adapters over FP16 and quantized weights

Multimodal AI • SmolVLM • LoRA

Vision-Language Data Pipeline

Pipeline converting structured driving-scene annotations into visual question-answer data for fine-tuning and evaluation.

  • Annotation-to-VQA data generation
  • Batched inference, LoRA fine-tuning, and exact-match evaluation

SLMs • Fine-Tuning • Evaluation

Small Language Model Reasoning

Experiment comparing prompting, supervised LoRA, and rejection fine-tuning for unit-conversion reasoning with SmolLM2.

  • Generated and filtered reasoning traces
  • Benchmarked numeric accuracy and parseable answers

Generative AI • Transformers • Quantization

Image Tokenization & Generation

Image-generation pipeline combining a patch autoencoder, binary spherical quantization, and an autoregressive visual-token transformer.

  • Patch encoding and discrete token mapping
  • Conditional and unconditional token generation

Mobile • SaaS • NFC

Perfect Synergy

NFC-powered digital business card platform with customizable profiles, instant contact sharing, and personal web destinations.

  • NFC tap-to-profile and contact download
  • Customizable digital identity workflows

Hobbies & Interests

Life beyond
the screen.

Recharging for me is exploring mountain trails, planning the next road trip, and finding offline balance with my dog, Marcus.

Beachside camp kitchen at sunrise.
campingBeachside camp kitchen at sunrise.
Desert highway pull-off after the rain.
road tripDesert highway pull-off after the rain.
Silhouette under the Milky Way.
astronomySilhouette under the Milky Way.
Trail overlook with Marcus.
hikingTrail overlook with Marcus.
Blood moon through the lens.
astronomyBlood moon through the lens.
Marcus cooling off in clear water.
hikingMarcus cooling off in clear water.
Misty lake framed by mountains.
photographyMisty lake framed by mountains.
Golden hour at the stadium.
photographyGolden hour at the stadium.
Big mountain views from the edge.
hikingBig mountain views from the edge.
Alpine lake beneath towering cliffs.
hikingAlpine lake beneath towering cliffs.
Storm clouds over a winding river.
photographyStorm clouds over a winding river.
Detailed moon captured against black sky.
astronomyDetailed moon captured against black sky.
Quiet beach walk with Marcus.
campingQuiet beach walk with Marcus.
Snowy mountains across a calm lake.
photographySnowy mountains across a calm lake.
Open road leading into the mountains.
road tripOpen road leading into the mountains.
Tent glowing beside the ocean at night.
campingTent glowing beside the ocean at night.