Ajayaditya L

Innovation Engineer (ProtoSem)

Forge Innovation & Ventures · Coimbatore, India

Jul 2025 - Jan 2026

Deep-tech startup incubator. Built and iterated deployable AI systems under compressed 2-week execution cycles, translating raw venture concepts into working architectures.

• Shipped 3 demo-ready AI prototypes (manufacturing QA, document intelligence, predictive maintenance) in 2-week build cycles at a government-backed incubator; one selected by an industry mentor for pilot-deployment evaluation.
• Designed citation-aware RAG pipelines combining dense vector search with BM25 lexical retrieval (hybrid scoring) for regulated verticals (healthcare, manufacturing) where every answer must trace to source documents.
• Optimized edge inference with quantized models (GGUF/ONNX) served via Docker + FastAPI, hitting sub-500 ms latency for 7B-parameter models on local GPUs; re-targeted real-time object detection to NVIDIA Jetson.

Applied AI / Full-Stack Engineer (Contract)

Trafy AI · Remote

Jan 2024 - Jun 2025

Multi-tenant AI-enabled B2B SaaS. Owned production LLM inference end-to-end, from database schema to the model routing layer, under live client workloads.

• Designed a multi-model routing layer (GPT-4, Claude, Gemini, Llama) with FAST/REASONING tiers selected on live cost, latency, and quality tradeoffs, cutting per-tenant inference cost 30%.
• Engineered a 4-tier provider fallback chain (primary → secondary → local quantized model → graceful degradation) with circuit breakers, cutting AI-feature downtime from ~12% to <0.5% and lifting reliability from ~88% to 99.5%.
• Implemented automated LLM-as-a-judge evaluation pipelines scoring groundedness, relevance, and correctness against human-validated rubrics; every model and router change was gated on these metrics before release.
• Drove a NoSQL → PostgreSQL re-architecture (row-level security) and a zero-downtime Firebase → Kinde migration enabling enterprise SSO/RBAC: query latency down 40%, 10x concurrent tenant load, zero critical auth incidents.
• Built the FastAPI REST layer with Celery/Redis async workers, tenant-isolation middleware, and rate limiting; shipped per-tenant cost, latency, and accuracy dashboards used by leadership to monitor AI system health.