Founder at APADCode (2025-09 – Present)
A B2B SaaS, AI-driven interview platform where agents conduct live-coding interviews to evaluate how candidates think, not what they output. Motivated from seeing the gaps in traditional screening tools and the rise of AI-assisted cheats, focusing on capturing deeper reasoning and communication signals through long-running, interactive AI interview. Built product from idea to a working B2B system within in 1 month by strategically leveraging AI tools - Coding assistants (Codex + Claude Code), strong context management with NIA, MCPs and custom Skills/Tools, etc. for fast iteration, while maintaining code consistency and system coherence.
- Architected the AI interview voice agent capable of conducting full-length (1 + hour long) live coding interviews with real-time interaction and evaluation, persistent context.
- Implemented low-latency conversational flow using LiveKit (deployed in-house) for streaming audio and managing bidirectional communication during interviews.
- Integrated ElevenLabs streaming TTS with chunked playback and latency control to keep responses natural without blocking the conversation loop.
- Engineered long-running context management by persisting structured state (Redis + MongoDB) for conversation history, code snapshots, and reasoning traces, with windowed context reconstruction to avoid token overflow.
- Built secure code execution using Docker-based sandboxes with per-session containers, CPU/memory limits, and timeout enforcement, ensuring safe and deterministic execution of untrusted code.
- Designed and implemented the cognitive evaluation layer that transforms live coding interviews into structured behavioral data for candidate assessment.
- Implemented a feature extraction pipeline in Python leveraging Claude APIs for semantic classification and custom deterministic logic for metric computation, generating features such as strategy latency, hint dependency, and probe alignment, etc.
- Designed scalable feature storage using MongoDB for structured metrics and Qdrant for semantic embeddings and retrieval of interview interactions.
- Developed feature-based evaluation and analysis workflows in Python (NumPy, Pandas) to enable consistent scoring and correlation of behavioral metrics with hiring outcomes.
- Architected Go portal backend services, implementing REST APIs with automated OpenAPI generation and enforcing RBAC via WorkOS through middleware-level authorization and tenant-aware access controls.
Founding Software Engineer at Practice AI (2023-08 – 2025-11)
AI-powered legal operations platform that automates client intake, document analysis, and case workflows for law firms. Built LLM-powered pipelines, RAG-based knowledge retrieval, and scalable backend services with modern, cutting-edge technologies. Ensuring strict compliance and secure handling of sensitive legal data was a core part of the system design from the day 1.
Defined the core AI system architecture from 0→1, building RAG pipelines, LLM orchestration, and end-to-end workflows while ensuring reliability in a high-stakes legal domain.
- Architected & led the development of multi-modal document parsing pipeline to process legal case data (PDFs, scanned medical records, images, and structured filings) into normalized, structured formats.
- Implemented OCR and layout-aware parsing using Tesseract, Azure Form Recognizer, PyMuPDF (PDF), and python-docx (DOCX), etc. with parser proxy logic to route documents by type and extract structured entities like injuries, timelines, and case metadata.
- Applied vision-language models (GPT-4o, later upgraded to GPT-5.4) to extract context from annotated records and handwritten notes beyond OCR, feeding outputs into structured parsing pipelines.
- Scaled ingestion and parsing with asynchronous pipelines (Celery, Azure Service Bus) and distributed Azure workers, orchestrating multi-stage workflows with task queues, retries, and backpressure control.
- Built compliance safeguards for sensitive legal data, handling PII and medical records with tenant isolation, encryption, and controlled data processing aligned with HIPAA and GDPR.
- Designed chunking strategies for legal documents, using layout-aware and semantic segmentation (sections, entities, timelines) to generate context-aware chunks optimized for embedding and retrieval.
- Developed AI-assisted document review dashboards using React, TanStack Query, React Hook Form, and Tailwind CSS, allowing attorneys to validate AI-extracted entities, review summaries, and approve automated case recommendations.
- Contributed to advanced data visualization interfaces using D3.js and Recharts, presenting case analytics, intake funnel metrics, and workflow automation insights to law firms through interactive dashboards.
- Championed adoption of AI coding assistants (Claude Code) across the engineering team, integrating them into daily development workflows and 5x boosting development productivity in feature development, debugging, and code iteration.
- Implemented multi-agent systems with LangChain/LangGraph to automate drafting, critique, and refinement cycles, improving document quality before human review.
Senior Software Engineer at DAS Technology (2021-04 – 2023-08)
As a full-stack engineer, worked on the Response Path CRM platform used by automotive dealerships to manage leads, automate messaging, and analyze customer engagement, owning end-to-end systems across real-time lead processing, workflow orchestration, multi-channel messaging, and backend platform unification in a high-scale, customer-facing environment. Worked closely with dealership clients to understand real operational workflows, adapt system behavior to their needs, and support live production usage, ensuring the platform aligned with how teams actually engage and convert customers.
- Played a key role in scaling the Response Path CRM platform by redesigning distributed backend services (Node.js, Python, Go) and modernizing frontend applications (React, TypeScript), enabling multi-tenant CRM workflows handling 1 M+ monthly customer interactions.
- Built the core lead response pipeline end-to-end, ingesting high-volume customer inquiries, enriching them with inventory and CRM data, and triggering personalized messaging flows across SMS, chat, and email using Node.js, Kafka, and PostgreSQL.
- Led backend integration and platform unification across multiple legacy services and acquired products.
- Replaced fragmented service-to-service calls with standardized REST/GraphQL APIs in Node.js and Python, aligning data contracts across teams.
- Built a unified customer data layer that resolved identities across CRM, dealer systems, and third-party sources, reducing inconsistencies in downstream workflows.
- Orchestrated cross-system workflows (lead → messaging → CRM sync) using event-driven pipelines with Kafka and RabbitMQ, eliminating duplicated logic across services.
- Developed real-time CRM analytics dashboards using React, Nivo, streaming APIs, and virtualized rendering, powered by Kafka streams and Redis caching to deliver near real-time operational insights to agents and marketing teams.
- Owned production reliability for customer-facing system — Response Path CRM —, operating in a forward-deployed model during remote rollouts.
- Built on-call runbooks and one-pagers documenting recurring failure modes (queue backlogs, retry loops, integration timeouts) to speed up incident response.
- Debugged real-time failures across lead ingestion and messaging pipelines (Kafka consumers, Node services, CRM sync issues), often under active customer usage.
- Acted as a direct point of contact for dealership clients, working closely with their teams to adapt workflows and resolve production issues aligned with their actual operations.
- Drove the migration from long-lived, manual release workflows to trunk-based delivery on AWS, setting up CI/CD pipelines with GitHub Actions, Terraform, and ArgoCD to enable continuous deployment and more reliable, incremental releases.
Software Engineer at Google (2018-06 – 2021-04)
Worked on Google's Contact Center AI (CCAI), building backend systems that turned conversational AI signals into real time, production workflows like CRM actions and Agent assist during live calls.
- Built backend services in Go and Java that translated Dialogflow NLP outputs into real-time actions - CRM updates, ticket creation, and call routing.
- Designed conversation state and workflow logic to turn unpredictable user input into deterministic, production-ready systems.
- Modeled multi-turn conversation state and translated NLP intent/entity outputs into structured backend actions like API calls, routing, and data updates.
- Added fallback and orchestration logic to handle ambiguity and edge cases, ensuring workflows stayed reliable even when AI signals were uncertain.
- Developed real-time pipelines for Agent Assist (a system that provides live recommendations and guidance to support agents during customer calls), processing live transcripts and serving in-call suggestions with low latency.
- Collaborated closely with ML engineers to productionize models via scalable APIs, using gRPC, Protobuf, and microservices on GKE with service-to-service communication and request routing.
- Improved system reliability and observability across microservices using GCP, Kubernetes, Redis, and distributed tracing (Dapper), metrics monitoring, and structured logging.
