ASR + NLP + Vision, pgVector, ANN, RAG, Behavioral Modeling

• Architected cross-modal pipeline (ASR + NLP + Vision, slot-filling, Gemini + pgVector ANN) over a 10K+ recipe / 30K+ alias database for daily meal logging, resolving dish matches in ~2s (DB-backed) to 3–5s (LLM fallback) at ~90% accuracy; powered a RAG meal-plan engine with behavioral signals and condition-specific diabetic constraints.

LoRA, QLoRA, GGUF, INT4, KV-Cache, Unsloth

• Led end-to-end VLM fine-tuning: 300K+ sample curation, quantity-aware supervision, and edge deployment with ≈68% compression (6 GB → 1.9 GB, INT4 GGUF) and ≤6s latency.

Vercel AI SDK, Kafka, HindSight, Redis, ClickHouse, Langfuse, OpenTelemetry, pgVector, Postgres

• Rebuilt AI Coach's agent orchestration layer — replacing an initial LangGraph state machine and dual-layer Mem0 memory with Vercel AI SDK's bounded tool-calling loop and a single HindSight episodic-memory layer — the foundation the meal-logging, diet-plan, and nudge systems below are built on.
• Built a 2-channel (WhatsApp, Telegram) conversational agent on a bounded (13-step max) tool-calling loop with a YAML-defined skill registry, driven by 16+ Kafka-consumed event domains (meal logging, glucose, onboarding) for real-time patient state.
• Built a single-call, 3-tier meal-logging resolution engine (personal history → pgVector recipe search → LLM fallback) with confidence-gated confirmation and cheap-first alternates sourcing, replacing a legacy multi-round-trip service; fixed production bugs in per-unit macro calculations and hardened LLM prompts against injection.
• Built an LLM-driven diet-plan generator (ClickHouse history + HindSight memory) and an 8-model behavioral reflection system (identity, routines, barriers, what-works), plus a 4-state glucose feedback engine closing the log → feedback → plan loop and driving personalized nudges.

• Contributed to SERB-funded behavioral biometrics project (deep learning, multilingual handwriting); processed 1200+ pages from 40+ authors for model training.