Secure SDKs and AI shopping agent built using AITP protocol

Delivered Typescript + Python SDKs that implement the new AITP protocol and a demo e-commerce agent that lets users shop Amazon directly from any MCP-compatible chat UI—built with clean architecture, 100 % test coverage, and penetration-tested for production-grade security.

Automated scalable AWS backend

Delivered infrastructure-as-code scripts to provision and auto-scale a microservice backend on AWS, including support for Trusted Execution Environments (TEEs)—a complex requirement rarely supported in cloud-native stacks.

AI-powered code generator that auto-deploys error-free code from specs

Created a multi-agent generator that ingests the AITP spec and outputs a fully functional, self-deploying agent codebase, automatically detecting and fixing type, compile, and runtime errors.

Ongoing end-to-end maintenance and security ops

Handled end-to-end maintenance, community triage, and vulnerability patching for Near’s open-source projects, ensuring long-term sustainability and developer trust.

AI-powered video analytics at scale

Built a computer vision system using CNNs and YOLO to identify truck types, capture license plates, and log entry/exit times from high-volume multi-camera streams—processing tens of terabytes daily across 10+ live construction sites.

AI-driven worker ID system enabling smart invoicing and access control

Delivered an AI module for worker identification that enabled contractor-specific invoicing and enforced geofenced access to site zones based on role or clearance.

AI-enabled PPE detection system automating safety compliance enforcement

Developed a mask and helmet detection system with smart alerting to ensure COVID-era and chemical site safety compliance, reducing manual oversight burden.

AI platform with scalable hybrid-cloud architecture and full lifecycle support‍

Designed for dynamic scalability (1000s of processing servers), hybrid cloud/on-premise processing, and maintained full lifecycle support for AI model training, deployment, and operations.

AI-powered secure access platform

Built a rule-based network security system with a natural language policy engine that learns from employee behavior and enforces real-time access control via a mobile-integrated platform, supporting sub-20ms latency at global scale.

Enterprise-grade backend architecture

‍Designed a distributed, service-oriented backend with multi-cloud and on-prem support, automated failovers, and >99.99% uptime to meet the needs of large-scale enterprise deployments.

AI-driven test scenario simulation

Developed an AI agent that generates and executes thousands of edge-case network test scenarios to ensure OpenVPN’s reliability under diverse and unpredictable conditions.

AI-enabled code generator and code reuse system

Delivered a monorepo architecture with shared utilities and AI-based code generators that convert DSL configs into platform-specific builds for iOS, Android, Windows, Linux, and MacOS—maximizing reuse and minimizing duplication.

LLM-powered multi-agent AI health coach

Built a personalized coaching system using LangGraph, orchestrating agents for fitness planning, nutrition, obstacle handling, and recipe recommendations. The system maintains evolving user profiles with both short- and long-term memory via vector embeddings.

Real-time transcription and insight extraction from voice calls

Developed a Gemini LLM-based pipeline to transcribe human coach calls and extract actionable insights, enabling real-time updates to user plans.

AI-driven recipe engine with Instagram integration

Delivered a personalized healthy recipe generator that connects to Instagram APIs, creates custom recipes from user inputs, and publishes SEO-optimized recipe webpages for increased engagement.

HIPAA and SOC-2 compliance, custom eval pipelines

Architected for millions of users with HIPAA and SOC-2 compliant infrastructure, integrated with leading EHR systems, and supported by dynamic prompt evaluation pipelines using LangSmith and Braintrust. Built with custom eval pipelines (LangSmith + Braintrust) to reduce hallucinations and ensure model quality through live and synthetic testing.

Scalable RAG AI answering dev queries from 100TB/day data

Built a RAG pipeline that ingests GitHub code, Linear tickets, Slack threads, specs, and more into Pinecone + Postgres, then powers an AI chatbot that answers how-to, debugging, and code-review queries—scaling to 100 TB/day with 1,000 + elastic workers and thousands of live webhooks.

LLM-powered Next.js library for real-time, context-aware engineering answers

‍Delivered a Next.js library that streams LLM responses and auto-renders the right component - plain text, code blocks, flowcharts, architecture diagrams, or performance charts - giving engineers rich, context-aware answers in real time.

Automated product planning from PRD to code with ticketing and reviews

Converted product requirements into draft design docs and populated Linear boards with ticket hierarchies, then checked pull requests for spec compliance—automating the planning loop from PRD to code.

SOC-2/HIPAA-compliant, 1-click AWS deployable stack & ongoing maintenance

Packaged the stack as Terraform for one-click AWS deployment, met SOC-2 & HIPAA via least-access controls and Vanta monitoring, and provided ongoing maintenance across multiple enterprise installations.

AI-powered process and workflow automation

Converted static process sheets into interactive web forms that adapt in real time to operator inputs, guiding staff through complex multi-step manufacturing workflows with dynamic instructions and parameter capture.

Digitization of handwritten process sheets

‍Built an OCR system using LlamaParse and Gemini multimodal models to turn pencil-written, smudge-prone factory notes into structured data with 90%+ accuracy, eliminating thousands of hours of manual entry.

Scalable industrial IoT data pipeline

Engineered a high-throughput system using AWS MQTT, Lambdas, and Clickhouse to process and store sensor data from thousands of devices.

Real-time smart alerting with AI agents

‍Implemented a LangGraph-based agent pipeline to trigger intelligent alerts from sensor data, identifying process anomalies and ensuring continuous monitoring across large-scale food and chemical manufacturing sites like Tyson Foods and Southern Chemicals.

Secure architecture & ongoing maintenance

Deployed secure VPN tunnels, fault-tolerant cloud/on-prem processing, and automated disaster-recovery workflows - now running at major food processors such as Tyson Foods - with Plank providing ongoing maintenance and support.

AI agent system to auto-create MLS listing videos in seconds

Built a multi-agent graph system that assembles MLS photos, clips, audio, and text into polished listing videos in < 5 s on GPU clusters, then auto-publishes to agents’ social  channels.

AI crawler unifying real estate data from Zillow, MLS, agents

‍
Developed an AI crawler that pulls images and metadata from agent sites, MLS, Zillow, and Redfin, cleans and deduplicates them, and unifies everything into a common schema for seamless editing.

AI tool for agents to auto-generate branded video content

‍
Delivered an AI-driven template framework letting agents swap fonts, backdrops, marketing copy, and property details to produce highly personalized videos that match their brand style.

Scalable AI system managing 1M+ listings, with ongoing maintenance

‍
Handles ~1 M listings with isolated databases for each brokerage, plus ongoing maintenance—already live with Sotheby’s agents.

Multi-agent graph-based AI agents analyzing financial data to deliver insights

‍
Built ~10 interoperating graph-based AI agents that automatically ingest, categorize, and analyze transactions from all connected financial institutions, surfacing tailored insights such as spending trends, budgeting nudges, and savings opportunities.

AI memory system for personalized financial advice from behavior data

‍
Developed a persistent vector-based memory system that captures financial behavior and user interactions, enabling agents to infer preferences, habits, and life context—powering deeply personalized financial advice.

AI-curated financial news feed

‍
Delivered an embeddings-powered indexing and ranking engine that classifies trending news and matches it to individual users' financial interests and holdings—creating a personalized Bloomberg-like dashboard.

AI platform with autoscaling, real-time inference, and regulatory security

‍
Designed a high-throughput system with dynamic autoscaling, secure hybrid deployment (cloud + on-prem), and real-time inference—paired with ongoing operational support and regulatory-grade data protection.

AI-powered document summarization and deal intelligence engine

Built a multi-agent system capable of summarizing massive document corpora (~10s of GB per deal) into a 128k-token context using a novel 3-step pipeline with SeTok (semantic vision tokenization), RAG-based chunking, and high-ratio compression.

Scalable pipeline for OCR across multiple data formats

‍OCR system supporting documents across heterogeneous file formats (PDFs, images, Excel), with SOC-2 compliance, least-access enforcement, idempotent data pipelines and robust error recovery.

Dynamic LLM calling to balance cost and answer quality

‍Implemented a dynamic token-usage optimization layer that selects the most efficient LLM-provider/prompt configuration per task, balancing cost and answer quality.

Eval datasets and  on-going prompt tuning

Deployed a Braintrust-based eval pipeline and trained prompt engineers to fine-tune domain-specific system prompts - automatically measuring precision/recall KPIs across evolving agent and model configurations.

AI-powered procurement research copilot

‍
Built a domain-specific AI agent system fine-tuned on battery chemistries, vendor catalogs, and robotics specs—enabling procurement teams to explore, compare, and source equipment with precision.

Autonomous vendor workflow automation

‍
Replaced manual procurement steps (email, fax, RFPs) with a multi-agent system that autonomously handles vendor inquiries, negotiations, and evaluations—learning from prior transactions to improve over time.

Scalable AI-powered crawler indexing terabytes of equipment data

‍
Engineered a smart crawling and indexing pipeline to fetch and deduplicate tens of terabytes of equipment data, ensuring a continuously updated knowledge base with minimal manual upkeep.

AI classification engine to auto-classify SKUs

‍
Developed an AI system that classifies equipment and SKUs using semantic embeddings and unsupervised topic modeling—creating a unified schema across diverse product types without hand-labeled categories.

Scalable, real-time AI pipeline to protect temperature-sensitive goods globally

‍
Deployed a real-time IoT data pipeline that ingests terabytes of temperature, location, and handling events, then uses event-driven AI agents to flag anomalies and push instant alerts—helping prevent the $30 B annual loss of temperature-sensitive goods.

On-device AI to autonomously maintain safe medicine temperatures in real time

‍
Designed an on-device AI algorithm that continuously tunes set-points in response to ambient conditions, removing the need for human intervention and ensuring medicines and vaccines stay within safe ranges in the field.

AI-powered compliance & audit engine with policy alerts, end-to-end traceability

Built a policy-based alert engine and automated, AI-driven audit report generator that produces secure, end-to-end traceability logs—dramatically cutting manual data-gathering time for regulatory reviews.

Secure data pipeline with ongoing maintenance, processing terabytes daily

Implemented encrypted network tunnels, Bluetooth/Wi-Fi auto-reconnect mobile apps, and fault-tolerant cloud infrastructure to process tens of terabytes per day, with Plank providing ongoing maintenance and uptime assurance.

AI-enabled reputation intelligence engine for tracking public perception

Built a fault-tolerant, horizontally scalable crawler that fetches millions of news, social posts, and SEC filings daily, then applies a hybrid evolutionary-algorithm + LLM pipeline to classify, summarize, and score company reputation across 50+ pillars like financial metrics, ESG, DEI and employee satisfaction.

LLM-powered news classification engine

‍Developed a multi-agent graph classification system using document embeddings and dynamic topic modeling to surface the most relevant articles for each reputation pillar in a web dashboard.

Optimizations to support large datasets for Fortune 500 companies

‍Implemented bandwidth-aware crawling, on-the-fly compression, and intelligent LLM invocation to cut storage and compute costs while supporting large datasets for Fortune 500 clients.

AI-driven chat UI that auto-generates live, explorable analytics

Delivered a chat-based analytics interface where user queries auto-generate the UI components that most effectively deliver insights for a given query and populates the compenents with live data using Text2SQL LLM agents, making insights deeply explorable.

AI engine to help businesses select optimal new store locations nationwide

‍
Built a multi-dimensional regression tree model trained on US census, cellular, and ordering data to score millions of intersections and recommend optimal store locations—used by major retail and restaurant chains nationwide.

High-performance AI-enabled geospatial visualization

‍
Delivered an SSE-based web map UI that streams server-rendered tiles and uses an AI model to reduce rendering load by intelligently narrowing the set of relevant intersections based on user input.

Real-time AI pipeline for nationwide sales and demographic analysis

‍
Engineered a concurrent data processing pipeline capable of re-running nationwide analyses in real time as users tweak parameters, handling massive datasets across geography, demographics, and sales history.

AI model orchestrator using LLMs to auto-select best algorithms

‍
Built a model orchestration system using LLM-judging to auto-select the best fit across regression, random forest, and graph-based models - also supporting ensembling for more accurate recommendations.

‍

Culturally adaptive AI voice copilot for healthcare

Built a HIPAA-compliant, voice-enabled assistant that supports real-time translation across Spanish dialects (Mexican, Haitian, Central American), enabling smooth communication between medical staff and non-native speakers through a fine-tuned OpenAI + Eleven Labs integration.

AI-powered voice forms that auto-translate and sync patient data to EMRs

Delivered a self-serve, web-based voice form that lets patients respond to questions in their native language—automatically translating, validating, and uploading data to EMR systems with minimal staff intervention.

AI-enabled OCR system that automates medical forms with 99% accuracy

Built an AI OCR pipeline with human-in-the-loop support that achieves 99%+ accuracy on complex handwritten medical forms, meeting strict HIPAA standards while drastically reducing manual workload.

Scalable AI system live in hospitals, reducing re-admissions and workload

Engineered for scale and reliability with seamless EMR integration, secure patient data handling, and ongoing maintenance—now live across multiple California healthcare facilities, reducing re-admissions and nurse burden.

AI-powered mobile app for real-time robotic fleet control

Built a Bluetooth-connected mobile app that fuses AI and manual control, enabling van operators to manage robotic systems in real time—handling both continuous sensor streams and bursty high-volume data, all within a secure local network.

Modular AI pipeline for in-van perception, planning, diagnostics

Helped design a modular, parallel-processing pipeline that routes in-van data to specialized AI models for perception (e.g., dimension/text recognition), motion planning, damage detection, packing optimization, and fleet diagnostics—deployable on both local hardware and cloud.

AI-driven UI for real-time delivery guidance and optimization

Delivered an adaptive UI component that provides real-time, AI-generated guidance to delivery personnel, optimizing package retrieval and handoff with dynamic instruction flows.

AI system for in-van/cloud use, piloting with global fleets

Engineered the system for dual deployment (in-van and cloud), with secure processing and resilient offline support—now in pilot phase for rollout with major global delivery fleets.

AI-driven energy usage prediction

Developed machine learning models that analyze historical and real-time data from the SPAN anel to predict future energy consumption patterns, enabling proactive optimizations that reduce bills by up to 20% through automated circuit adjustments.

Intelligent backup power allocation

Implemented AI algorithms using reinforcement learning to dynamically allocate backup power during outages, prioritizing critical circuits based on learned household behaviors and integrating with EV charging via SPAN Drive for seamless, efficient resource distribution.

LLM-powered insight generation

Created a natural language processing system within the SPAN home app that generates personalized, human-readable energy reports and recommendations, contextualizing appliance-level data with user queries for intuitive energy management.