Experience

What I Owned

Everything. An AI voice-agent platform for car dealerships built end-to-end solo - telephony, speech pipeline, CRM, instrumentation, and the customer relationships. Live in production on a paid pilot with the world's largest Hyundai dealer group (62 outlets), with a $60K committed annual pipeline across four dealer groups.

What I Built

• Per-call tracing end to end - a timeline from the WebSocket accept through first audio to call end, per-turn perceived latency attributed to STT, LLM and TTS, phantom-turn suppression, and unit cost measured at $0.02/min from production, not vendor price sheets. A total with a missing component reports null, never a made-up number
• Real-time voice pipeline (Python, FastAPI, Pipecat, WebSockets) over 8kHz telephony streams - 13 Indian languages including code-mixed Hinglish, swappable STT/LLM/TTS engines (Sarvam, Gemini, ElevenLabs), 300ms p50 / 490ms p95 speech-to-speech reply latency
• Delivery rules for telephony and WhatsApp - the recording URL arrives minutes after the call on a separate callback and is matched by call ID; a messaging or recording failure can never fail the call record; an ambiguous phone number means do not message, never message and hope
• Dealer CRM (Next.js, TypeScript, Drizzle, PostgreSQL) with transcripts, lead capture, and WhatsApp staff alerting through the dealer's existing provider
• Dual-service GCP architecture - GCE VM for long-lived call WebSockets, Cloud Run for the CRM - with CI/CD that drains live calls before deploys and gates on generated DB migrations

Interesting Problems

• An unpinned dependency took production down: a rebuild resolved the voice framework to a newer version whose STT model list had changed, and every inbound call died at construction. What changed after: the pin, a pre-bump checklist (the provider's model allow-list, the serializer patch), and the rule written where the next engineer will read it
• The first caller after every restart heard silence: a 12-second cold import sat inside the call path. Heavy imports and credential fetches now live in the server's startup hook, and measuring the timeline before guessing is the rule - two confident latency hypotheses were refuted in under a minute each
• Cut per-turn latency by 1.2s: the STT framework knew transcripts were final but never set the finalized flag, so every turn silently waited out the provider's full timeout
• My latency dashboard lied in both directions - phantom turns inflated p95 from 490ms to 2.5s while the clock started 200ms late - so I rebuilt the measurement before touching the pipeline
• Ran the full customer lifecycle solo: interviewed 15-20 of Bangalore's largest dealer groups, demoed to owners and GMs, closed the paid pilot, and iterated agents from call recordings and staff feedback

What I Owned

End-to-end ownership of the core campaign and incentives platform at a Web3 growth company with 10M+ users - distributed backends, a customer data and experience platform, identity verification infrastructure, an AI website editor, and high-throughput data pipelines powering enterprise Web3 engagement programs.

What I Built

• Distributed ingestion systems processing millions of user submissions with concurrency-safe pipelines, rate limiting, and moderation workflows that handled traffic spikes without breaking
• A CDXP end-to-end - funnel event tracking, drop-off detection, cohort segmentation, and dashboards across ~2M user events that surface exactly where users fall out of each campaign
• AI-powered website editor with natural-language edits, live preview, and structured diff commits - a non-engineer can ship a branded landing page and iterate on it without touching code
• dbt transformation models powering campaign points and incentive systems - large-scale data cleanup and pipeline failure resolution across millions of rows of user activity data
• Reputation-based identity verification system using non-transferable NFTs

Interesting Problems

• Reduced infrastructure costs from $9,000/month to $500/month (94% reduction) by moving a server-rendered Next.js app off AWS Amplify onto 8-9 containers on a private VPS behind Caddy - a route-by-route split of server, client and static, a shared Redis cache replacing per-container caches to fix coherence, release-tied invalidation with rolling restarts, and a DNS cutover under live traffic
• Designing a funnel event schema flexible enough to track every user action across campaigns but stable enough that dashboards and downstream analytics do not break every time a new campaign ships
• Debugged distributed cache inconsistencies across production workloads - the kind of bug where everything looks fine on paper but users see stale data
• Optimized inbound/outbound APIs and database queries to eliminate bottlenecks in a system serving enterprise-scale traffic

What I Owned

Cloud migration of legacy on-prem trading systems and real-time data synchronization infrastructure for 350k+ active traders.

What I Built

• Migrated on-prem Spring Boot systems to AWS using Terraform - Lambda, DynamoDB, S3, SES - achieving 35% cost reduction with 99.99% uptime and sub-50ms query latency
• Distributed Kafka pipelines and WebSocket services synchronizing real-time trading data between cloud and on-prem systems, reducing manual data handoffs by 80%
• Spring Boot microservices on Nomad processing 10k+ Kafka events daily with zero downtime, including integrations with ProRealTime and TradingView

Interesting Problems

• Building reliable event processing at scale where a dropped message means a trader sees wrong positions
• Designing the cloud↔on-prem sync layer where eventual consistency meets financial data that needs to be exact
• Third-party trading platform integrations where APIs are poorly documented and failure modes are discovered in production

What I Owned

Architecture and engineering for a multi-tenant apprenticeship platform serving African markets. Everything from database design to payment integration to real-time messaging.

What I Built

• Multi-tenant backend with Supabase RLS, PostgreSQL triggers, and serverless functions enforcing fine-grained access control across mentors, apprentices, and organizations
• Real-time messaging system using Supabase Realtime and PostgreSQL LISTEN/NOTIFY for low-latency cohort communication
• Paystack Split Payments integration automating apprentice-to-mentor payouts with compliant fee extraction

Interesting Problems

• Onboarding users with varying technical literacy - 12 iterations to reach 70% completion rate
• Payment compliance across different African markets with different regulatory requirements
• Row-level security policies that are simple enough to audit but flexible enough to handle complex org hierarchies

What I Owned

Contributed to a real-time logistics tracking platform used by manufacturers and transporters - my first production code.

What I Built

• React/Next.js components and Node.js REST endpoints for live order tracking, with Google Maps integration and WebSocket-powered vehicle location updates
• An in-browser PDF and image viewer using React and PDF.js for cross-device receipt rendering

Interesting Problems

• Keeping live location updates usable over intermittent connectivity in rural India