Pratham Adhikari

AI Engineer · Kathmandu, Nepal

• Owned the AI integration layer across the product, from spec to CI/CD to production observability, integrating Google Gemini APIs and self-hosted LLM, ASR and TTS services into customer-facing backend systems.
• Designed and shipped end-to-end RAG pipelines over heterogeneous corpora — US tax and regulatory documents, and crawled web content — handling ingestion, semantic and recursive chunking, embedding generation, and vector store selection across pgvector and Solr; benchmarked chunking strategies and tuned retrieval parameters against evaluation sets to improve answer relevance.
• Currently architecting the retrieval layer to scale from thousands to millions of documents, evaluating index partitioning, embedding throughput and cost-per-query tradeoffs across vector store options.
• Built agentic and conversational workflows using LangChain, LlamaIndex and Pipecat, including tool-call design, context aggregators and custom orchestration for multi-step flows.
• Established the team's prompt engineering practice: system prompts, few-shot and chain-of-thought patterns, structured output schemas, config-driven prompt versioning, and regression evaluation suites in Promptfoo to catch quality drops before release.
• Implemented user-level style and preference personalisation: typed preference schemas persisted in the database and injected into system prompt construction at request time, giving consistent per-user tone and response style across sessions without retraining.
• Implemented LLM output guardrails in Pipecat — schema-based validation, intent-based routing, branched content filtering and fallback logic — applied to both chat and real-time voice-to-voice agents to prevent unsafe or malformed responses reaching users.
• Instrumented production AI observability with Langfuse, Grafana, MLflow and Dozzle, tracking latency, token usage, cost-per-query and output quality drift to drive optimisation decisions.
• Architected real-time voice-to-voice (V2V) agents orchestrating ASR, LLM and TTS microservices over WebSockets, achieving sub-3-second end-to-end response latency for English and sub-4-second for multilingual conversation through streaming TTS delivery, jitter buffering and pipeline-stage parallelisation.
• Engineered backend microservices in Python/FastAPI with ZeroMQ message-queue orchestration and Docker deployment; implemented authentication and authorisation securing internal APIs.
• Built telephony call infrastructure on Asterisk PBX — multi-language IVR flows and automated inbound/outbound AI-driven call campaigns; load-tested to a fully saturated 16-channel SIP trunk with concurrent live AI agents holding latency targets.
• Fine-tuned Whisper for Nepali and code-mixed ASR, reaching 33% WER on a low-resource language and delivering the organisation's best-performing model; owned the full data pipeline (collection, annotation, filtering, preprocessing). Also fine-tuned TTS and translation models (Chatterbox, Parler-TTS).
• Deployed and maintained on-premise model serving and inference infrastructure alongside Google Cloud services, ensuring data privacy and low-latency performance.
• Led the speech research and backend teams — setting research direction, running code reviews and coordinating delivery across both groups; collaborated with frontend engineers to surface AI capabilities in product UIs.