AI ENGINEER ↗
A little logic. A lot of curiosity.I’m Ninad. I build at the intersection of AI, code, and human curiosity. Thoughtful systems. Tangible experiences.
EXPLORE MY WORK ↓PERSONAL PORTFOLIOA LITTLE CONTEXT / 01
I like the part where an idea stops being abstract and starts doing something. That’s what draws me to AI: the room to experiment, the hard questions, and the craft of making it useful.
WHAT I WORK WITH ↗engineer = { "name": "Ninad Naik", "focus": "applied AI", "mode": "always learning" }
EXPERIENCE / 02
Applied AI engineering across LLMs, data systems, and multimodal product workflows.
Germany — Remote · Jul 2026 – Present
Building LangGraph conversational flows, FastAPI services, and RAG pipelines for a reliable AI interview platform. I also design the PostgreSQL/Supabase layer for generated assets, validation, indexing, access policies, and lifecycle handling.
US — Remote · May 2026 – Jul 2026
Worked on creator-facing AI pipelines that coordinate speech, language models, memory, and media generation. Built Playwright and FFmpeg automation, and integrated ElevenLabs and Tavus for voice and video workflows.
Feb 2026 – Mar 2026
Developed representation-learning and anomaly-detection experiments over 50K+ records. Used Autoencoders, PCA, UMAP, threshold selection, and validation to reduce false-positive detections by approximately 30%.
Hyderabad, Telangana, India · Remote · Mar 2026 – Apr 2026
Worked on transformer inference and causal behaviour in LLMs, using controlled interventions and token-level analysis to study how information propagates during generation. Focused on experimental design, evaluation methodology, and analysing model behaviour without modifying the underlying model weights.
INDEPENDENT RESEARCH / 03 · IN PROGRESS
I’m testing whether every previous token matters equally during autoregressive generation. The framework changes individual token contributions in frozen LLMs through KV-cache ablation, resampling, and noise injection, then measures future generation with matched-seed paired rollouts.
The aim is to compare attention, output divergence, and acceptance-rate proxies against direct causal measurements across open 8B-scale models.
If causal dependence concentrates in a small set of positions, it could inform KV-cache usage, speculative decoding, and inference-time computation.
SELECTED WORK / 02
Research systems, evaluation tools,
and independent experiments.
Forensic computer vision
A paper-inspired system for distinguishing genuine ID captures from print and screen presentation attacks.
Synthetic document intelligence
A reproducible VLM and OCR benchmark for structured extraction, tamper detection, calibration, and failure analysis.
Autonomous Mars EDL
An autonomous guidance research system for simulated Mars entry, descent, and powered touchdown.
LLM reliability research
An evaluation environment for measuring whether uncertainty signals predict factual errors in LLM outputs.
HOW I THINK / 03
Understand what needs to work, who it’s for, and what success actually looks like.
Build a small version. Test the assumptions. Learn from what breaks.
A good system should be useful, understandable, and a pleasure to interact with.
THE WORKING STACK / 04
The landscape I like to build in.
GOOD THINGS START WITH A CONVERSATION.
AI engineer building research systems,
evaluation tools, and useful products.