Akshath Yennam

WACC Compiler

A full compiler for WACC (a C-like language) in Scala, from lexing through AArch64 codegen, extended with a CFG-based optimisation pipeline (constant propagation, dead code elimination) and a graph-colouring register allocator with multiple spill heuristics.

• +Implemented data lineage emission across ~100 Investment Analytics ETL pipelines, wiring them into the firm's DataHub platform and replacing bespoke per-pipeline code with a reusable onboarding pattern.
• +Built declarative data-quality checks (freshness, completeness, statistical anomaly detection) running at every ETL load step, catching issues before they reach client-facing dashboards.
• +Folded lineage and data-quality into a config-driven pipeline base class, so future pipelines inherit both by default with no extra setup.
• +Architected the system event-driven on Kafka, with checks executed on Ray clusters inside an isolated Kubernetes namespace.

• +Architected and extracted a reusable omnichannel customer-support platform core from a production system, with an overlay layer letting client-specific workflows build against one core interface without forking it.
• +Designed and deployed a scalable omnichannel memory layer for real-time AI agents in a live Samsung Mexico deployment, gating model downgrades on eval results to cut inference cost 30% with no regression.
• +Cut live response latency 70% by moving memory extraction and knowledge-graph (Neo4j) writes off the synchronous path into an async job queue.
• +Built an evaluation and observability layer (prompt versioning, regression testing, per-agent cost tracking), taking the system from prototype to a production-ready v1 used in real customer interactions.

• +Rebuilt a JD-to-resume semantic search system on the Qdrant vector database, cutting query response times from ~90s to ~6s — a 15x speed-up — through caching and search-parameter tuning.
• +Built a hybrid retrieval pipeline combining vector and keyword search with multi-query rewriting and reciprocal rank fusion, improving retrieval quality for a sales-meeting assistant.
• +Built a real-time voice interviewing agent on GPT-Realtime with a multi-agent, phase-by-phase architecture for context-aware handoffs.
• +Implemented prompt engineering and guardrails to keep live interviews on-topic, on-tone, and hallucination-free.

• +Got an early introduction to machine learning, training basic regression and clustering models.
• +Explored transformers and NLP, building a sentiment-analysis project and a text-classification project on unlabelled data.
• +Took a first look at computer vision — learned how CNNs work and built a simple image classifier.