Eight hundredmilliseconds.

Built a RAG-powered conversational analytics agent with LangChain, pgvector, and OpenAI embeddings, so non-technical operators can query account-specific industrial datasets in plain English. Multi-stage async pipeline: semantic retrieval over a vector store, NLP-driven dynamic SQL generation, multi-turn context memory, structured output validation.

95% query accuracy · 60% less analyst workflow time

Designed a tool-use agent framework on LangChain tool-calling with configurable retrieval chains, plus an LLM eval harness benchmarking retrieval precision and SQL execution correctness. Prompt iteration and model upgrades ship on evidence now, not intuition.

200-query golden dataset

Replaced the flat 2D machine schematic with an interactive Three.js digital twin that mirrors the real asset: the train built from real component bodies, sensors drawn at the mount points they are fitted to, and live vibration, temperature and open alarms rendered onto the geometry itself, so a bearing going bad shows up on that bearing rather than in a table beside it. Sensors are fitted by dragging onto a mount point and the train is edited in the scene, with layout changes applied optimistically and rolled back on refusal, asset model invariants enforced in the interaction rather than returned as a 400, and a 2D fallback where WebGL is unavailable.

75 component types modelled

Owned a high-availability Annotations System across microservices with Redis caching and RBAC, with full audit logging on every write path.

sub-second responses · >99% uptime

Worked directly with reliability engineers and plant operators to turn ambiguous maintenance workflows into product requirements, then shipped against them end to end.

Owned the full pipeline for production LLM voice agents on live phone calls: telephony media ingest, streaming ASR, intent resolution, TTS playback back onto the live call, all designed against a sub-800ms turn-latency budget.

sub-800ms turn budget

Solved the hard parts of live conversation rather than the happy path: barge-in with in-flight TTS cancellation, endpointing and turn detection on partial ASR results, and graceful degradation when any single stage slowed. A slow LLM call never blocked the audio loop.

Engineered a real-time call intelligence pipeline in Node.js, Django, Redis, and Whisper, turning thousands of call recordings into transcripts, multi-label sentiment scoring, and agent performance summaries.

+20% analysis accuracy · 6 hrs/analyst/week of manual QA removed

Designed an event-driven ingestion system on Apache Kafka and Node.js consuming call lifecycle events at scale. Exactly-once semantics anchored downstream through idempotent writes and unique constraints, rather than trusting the transport layer.

40% lower event processing latency · zero data loss at peak

Shipped a context-aware LLM co-pilot for live agents on GPT-4 with RAG over internal knowledge bases: in-call suggestions, next-best-action prompts, dynamic FAQ resolution.

+25% CSAT across 3 enterprise customers

Architected the Automated Contact Center platform frontend in TypeScript with Next.js and a Turborepo monorepo, introducing code-splitting and edge caching.

30% faster page load · +25% engagement

Built the video layer for LiveChat on Django, WebSockets, and the Dyte SDK: meetings created and participant tokens minted server-side so no call credential ever reached the browser, with signalling and presence riding the WebSocket channel the chat already held open. The work that mattered was the failure paths rather than the happy path: denied camera and microphone permissions, rooms left orphaned by a closed tab, and a mid-call network drop that has to put the customer back into the same session instead of a fresh one.

30% fewer video support tickets · +18% session engagement

Rebuilt LiveChat agent assignment from a polling scheduler into an event-driven service on Django, Celery, and Redis. The old path swept every 30 seconds, scanning every waiting customer and matching each to a free agent by query domain and preferred language, so a request raised just after a sweep waited out the whole interval, and the sweep still ran on the cycles where nothing was waiting. Assignment now fires off the request event itself: routed on domain and language against live agent availability, with an atomic claim so two concurrent events cannot take the same agent, and bounded retry with backoff that re-queues a request rather than spinning when no match is free.

30s worst-case wait removed · 110 hrs/week of manual ops removed

Brought Instagram and Facebook into LiveChat Fusion as first-class channels on the Meta Graph API, normalising each platform's payload shape into one conversation model so an agent worked a single queue instead of a tab per network. Inbound webhooks acknowledge immediately and fan out to workers, because Meta redelivers anything it does not get a fast 200: messages are deduplicated on the platform message ID so a retry cannot post twice, and outbound sends are shaped to the per-page rate limit and the messaging window rather than finding them at runtime.

3 messaging channels · 15% faster average response

Built and tested REST APIs for communication between application components, working to the reliability bar of a platform that was already in production.

Debugged and resolved live production issues, which is where I learned to read a stack trace and a log timeline before touching any code.

Worked the full cycle inside an agile team: code review, design discussion, and feature work shipped to deadline.