Bikash Lama

Grew a small MVP into a multi-tenant SaaS over two years of continuous delivery — now integrated into an ecosystem where the same restaurant is discoverable and reservable across collaborating partner platforms.

• Google reservation flow + cross-platform listing — one restaurant is bookable from any partner in the ecosystem.
• AI-in-the-loop optimization: used Claude and Cursor to surface API bottlenecks and tighten hot-path UI renders.
• Multi-OEM architecture deployed via Terraform with a dev → staging → production CI/CD gate.

An internal tool that translates any Japanese document — Excel, Word, drawio diagrams — to English via Claude on the fly.

• Format-preserving extraction across xlsx, docx, drawio — formulas, styles, and shape positions stay intact; only the human-readable text is replaced.
• Translation cache keyed on normalized source text + target language: identical strings are never paid for twice.
• Tiered model routing — bulk and low-complexity content goes through cheaper models; technical or mixed-domain text routes to the larger one.

A WebRTC-based real-time consultation product that lets customers talk live to agents across sectors — banking advisory, telemedicine, retail product demos. COVID pushed demand sharply, forcing a hard-deadline reengineering of the thin MVP into a production-grade architecture without slowing the delivery cadence.

• Real-time video consultation over WebRTC + Twilio, shipped across banking advisors, online clinics, and apparel retail under one product surface.
• Re-engineered the MVP skeleton from the ground up as customer load grew — moved from 'demo works' to '100% uptime SLA' shape.
• Observability pipeline tuned for live-call triage: structured logs make it possible to chase a dropped call back to the exact request that failed.

A chatbot for internal document Q&A. The LLM writes the answer; my code checks every citation before the user sees it. Ungrounded answers get replaced with a refusal server-side.

• Streamed answers get replaced with a refusal if a parser can't tie every [N] citation to a real retrieved chunk.
• Refusal short-circuits before the LLM even runs when retrieval returns nothing above threshold — no tokens spent, no hallucinated citations.
• Per-user daily token caps enforced in application code. Every OpenAI call writes cost to a `usage_events` table for the admin dashboard.

A Python script that drives a real browser via plain-English instructions. First draft was 80 lines and lied about everything. Rebuilding it produced a reusable pattern I now apply to every piece of LLM-touching code.

• Every tool call is validated by code; success is proven by re-reading state, not by the LLM's claim.
• Cost and iteration ceilings live in the harness — the LLM can't loop indefinitely on a broken action.
• Refusals short-circuit in code, not by prompt persuasion. Same pattern powers the RAG assistant's citation check.